Eigen-Based Signal Processing Methods for Ultrasound Color Flow ...
Eigen-Based Signal Processing Methods for Ultrasound Color Flow ...
Eigen-Based Signal Processing Methods for Ultrasound Color Flow ...
Create successful ePaper yourself
Turn your PDF publications into a flip-book with our unique Google optimized e-Paper software.
<strong>Eigen</strong>-<strong>Based</strong> <strong>Signal</strong> <strong>Processing</strong> <strong>Methods</strong> <strong>for</strong> <strong>Ultrasound</strong><br />
<strong>Color</strong> <strong>Flow</strong> Imaging<br />
by<br />
Alfred C.H. Yu<br />
A thesis submitted in con<strong>for</strong>mity with the requirements<br />
<strong>for</strong> the degree of Doctor of Philosophy<br />
The Edward S. Rogers Sr. Department of Electrical and Computer Engineering<br />
Collaborative Program with the Institute of Biomaterials & Biomedical<br />
Engineering<br />
University of Toronto<br />
© Copyright by Alfred C.H. Yu 2007
ABSTRACT<br />
<strong>Eigen</strong>-<strong>Based</strong> <strong>Signal</strong> <strong>Processing</strong> <strong>Methods</strong> <strong>for</strong> <strong>Ultrasound</strong><br />
<strong>Color</strong> <strong>Flow</strong> Imaging<br />
Alfred C.H. Yu, 2007<br />
Edward S. Rogers. Sr. Department of Electrical and Computer Engineering &<br />
Institute of Biomaterials and Biomedical Engineering,<br />
University of Toronto,<br />
Toronto, ON, Canada<br />
Purpose: This thesis presents a study on the design of new eigen-based signal processing<br />
methods <strong>for</strong> use in color flow imaging. Specifically, the proposed methods are designed to<br />
address three main problems in color flow signal processing: the lack of abundant Doppler data<br />
samples, the possible presence of wideband Doppler clutter, and the potential flow signal<br />
distortions that arise during clutter suppression.<br />
Theoretical Contributions: A clutter filter (the Hankel-SVD filter) and a flow estimator (the<br />
Matrix Pencil method) were respectively developed by exploiting the eigen-space principles<br />
related to two matrix <strong>for</strong>ms known as the Hankel matrix and the matrix pencil. Both techniques<br />
were then combined to <strong>for</strong>m a new color flow data processor that per<strong>for</strong>ms flow detection and<br />
flow estimation in parallel. All of these methods are adaptive to the Doppler signal contents<br />
through an SVD of the Hankel matrix created from a Doppler data vector. Moreover, they are<br />
intended to work with each Doppler ensemble separately (i.e. they do not require Doppler<br />
ensembles from multiple sample volumes).<br />
Simulation Assessment: A Doppler signal simulation model was developed to generate Doppler<br />
signals consisting of non-stationary (phase-modulated) tissue clutter and spectral-broadened<br />
(amplitude-modulated) flow echoes. The synthesized datasets were used to analyze the Hankel-<br />
- ii -
Abstract<br />
iii<br />
SVD filter’s flow detection per<strong>for</strong>mance and the Matrix Pencil’s velocity estimation<br />
per<strong>for</strong>mance. A comparison of the Hankel-SVD filter with an existing type of adaptive filter<br />
(clutter-downmixing filter) showed that this new filter is more capable of discriminating flow<br />
echoes from Doppler clutter. Also, it was found that the Matrix Pencil can provide less biased<br />
flow estimates as compared to other frequency-based flow estimators like the lag-one<br />
autocorrelator.<br />
Experimental Assessment: <strong>Color</strong> flow datasets were obtained respectively from a steady-flow<br />
phantom and from the carotid arteries of a youth subject. These datasets were used to analyze<br />
the per<strong>for</strong>mance of the proposed single-module processor. For the flow phantom study, the<br />
single-module processor showed that it can reconstruct velocity maps similar to the theoretical<br />
flow profile. As well, <strong>for</strong> the in vivo studies, it gave a better flow detection per<strong>for</strong>mance than<br />
conventional two-module processors especially when there is substantial tissue motion.
ACKNOWLEDGEMENTS<br />
It was a snowy winter day when I first officially started my doctoral study at the<br />
University of Toronto. From then to now, this research journey has been rather overwhelming<br />
with many rough rides in between. In particular, many long school days were spent on the 4 th<br />
floor of Rosebrugh Building and the 3 rd floor of Mining Building as I tried to decipher various<br />
abstract concepts from the literature. Throughout the course of study, many important<br />
individuals have offered advices and support that are invaluable to me, and it is my honor to<br />
receive their words of wisdom. Prior to the presentation of this dissertation, I would like to give<br />
big credit to these important individuals.<br />
First and <strong>for</strong>emost, I would like to express my sincere gratitude to my thesis advisor, Prof.<br />
Richard Cobbold, <strong>for</strong> giving me the freedom to explore and develop my own research directions.<br />
Dr. Cobbold has always been very supportive of my work and has provided lot of insightful<br />
comments to help me fine-tune my research results. My co-advisor, Dr. Wayne Johnston, has<br />
also been an influential figure throughout my study. I have really appreciated his strong<br />
encouragement along with his great sense of humor.<br />
During a random occasion, I had the chance to meet with one of our lab’s alumnus, Dr.<br />
Larry Mo. After our initial discussion on that day, we eventually turned out to become<br />
collaborators in research. I would like to thank him <strong>for</strong> sharing his insights on the design aspects<br />
of ultrasound systems. I am also very grateful <strong>for</strong> his help in acquiring parts of the experimental<br />
data used in this thesis.<br />
When I first started my doctoral study, Prof. Raviraj Adve introduced me to the concept<br />
of eigen-based signal processing. I really appreciated his advice and insights in this research<br />
field that eventually evolved as the theme of this thesis. Special thanks also go to Prof. Michael<br />
Joy and Prof. Berj Bardakjian <strong>for</strong> their ever challenging questions on the fundamental<br />
motivations of my studies. Their constructive comments have stimulated many after-thoughts<br />
following every committee meeting and defense.<br />
The occasional visits to the 6 th floor of the Research Wing at Sunnybrook & Women’s<br />
College Health Sciences Centre were always fun and interesting. I would like to thank Prof.<br />
- iv -
Acknowledgements<br />
v<br />
Peter Burns <strong>for</strong> granting me access to the ultrasound scanners in his laboratory. I also wish to<br />
thank Dr. Milan Banerjee and other members of the Burns research group <strong>for</strong> helping me resolve<br />
various problems encountered during the data acquisition process. It is very un<strong>for</strong>tunate that Dr.<br />
Banerjee passed away be<strong>for</strong>e this thesis was completed, but his smiles and humor will always be<br />
remembered.<br />
This research would not have been completed without the help of Dr. Matthew Bruce<br />
from Philips <strong>Ultrasound</strong>. I am grateful <strong>for</strong> his valuable help on fixing the Philips HDI-5000<br />
machine that I used to collect parts of my experimental data. I would also like to thank Dr. Eric<br />
Cohen-Solal from Philips Research <strong>for</strong> offering me a special opportunity to work as a research<br />
intern and learn about the developmental aspects of medical ultrasound systems. This internship<br />
experience has widely broadened my knowledge in the ultrasound field.<br />
School has always been a mix of pain and pleasure. Although the quest <strong>for</strong> knowledge<br />
has often been a struggle, the joy of working alongside with the past and present members of the<br />
Cobbold research group has certainly been a pleasant experience. I will also remember those<br />
Dilbert-style chats with officemates at RS 411 and classmates from many courses.<br />
Perhaps the most gracious moment at the end of a long school day was the subway ride<br />
back to my uptown home. During the ride from Queen’s Park to Finch, I usually collapsed into a<br />
state of coma, with anything happening around me being considered as white noise. The<br />
moment when I opened the door of my home was ever pleasing, as those familiar voices were<br />
often there yapping about my unpunctuality <strong>for</strong> supper. I would like to thank my parents,<br />
Frankie and Candy Yu, <strong>for</strong> moving over to Toronto during my last year of study. With their<br />
great backline support, I never had to worry about my dietary needs. It is also my honor to<br />
receive strong encouragements from my <strong>for</strong>mer housemates Simon Chan and Ka-Wai Leung.<br />
They have always put up a plat<strong>for</strong>m <strong>for</strong> me to unload my personal frustrations. As well, I have<br />
really appreciated the occasional morale boasters from all my great friends: especially the<br />
entertaining ones from Martin Mak, Jenny Choi, and fellow Yu’s Club members.<br />
Lastly, I would like to acknowledge the scholarship support from the Natural Sciences<br />
and Engineering Research Council of Canada as well as the short-term funding from the Walter<br />
Sumner Foundation and the Fraser Elliot Chair of Vascular Surgery. Their generous support has<br />
greatly assisted my financing throughout this thesis study.
TABLE OF CONTENTS<br />
ABSTRACT ................................................................................................................................. ii<br />
ACKNOWLEDGEMENTS.................................................................................................................. iv<br />
TABLE OF CONTENTS .................................................................................................................... vi<br />
LIST OF ABBREVIATIONS AND SYMBOLS ....................................................................................... x<br />
CHAPTER 1 Introduction to <strong>Color</strong> <strong>Flow</strong> Imaging................................................................... 1<br />
1.1 Clinical Background ............................................................................................................. 1<br />
1.2 Outline of Thesis Study ........................................................................................................ 2<br />
1.2.1 Research Motivation and Hypothesis ............................................................................ 2<br />
1.2.2 Goals and Objectives ..................................................................................................... 2<br />
1.2.3 Chapter Overview .......................................................................................................... 3<br />
1.3 Overview of Doppler <strong>Ultrasound</strong>.......................................................................................... 4<br />
1.3.1 General Concepts........................................................................................................... 4<br />
Physical Principles ............................................................................................................. 4<br />
Doppler Equation................................................................................................................ 4<br />
Range Localization ............................................................................................................. 5<br />
1.3.2 Display Modes ............................................................................................................... 6<br />
Spectrogram........................................................................................................................ 6<br />
<strong>Color</strong> <strong>Flow</strong> Image............................................................................................................... 7<br />
M-Mode Profile................................................................................................................... 7<br />
1.4 Principles of <strong>Ultrasound</strong> <strong>Color</strong> <strong>Flow</strong> Imaging...................................................................... 8<br />
1.4.1 System-Level Overview .................................................................................................. 8<br />
Data Acquisition ................................................................................................................. 8<br />
<strong>Signal</strong> <strong>Processing</strong>................................................................................................................ 9<br />
1.4.2 Imaging Considerations............................................................................................... 10<br />
Transducer Type and Transmit Pulse Characteristics ..................................................... 10<br />
Image Resolution and Penetration Depth......................................................................... 12<br />
Velocity Resolution and Aliasing Limit............................................................................. 13<br />
Pulse Repetition Interval and Doppler Ensemble Size ..................................................... 14<br />
Parameter Optimization.................................................................................................... 15<br />
1.4.3 Advanced <strong>Flow</strong> Imaging Approaches .......................................................................... 15<br />
Three-Dimensional Mapping............................................................................................ 15<br />
Vector <strong>Flow</strong> Imaging ........................................................................................................ 16<br />
High-Frequency <strong>Flow</strong> Imaging......................................................................................... 17<br />
Microbubble Contrast Agents........................................................................................... 17<br />
Additional Remarks........................................................................................................... 18<br />
1.5 <strong>Signal</strong> <strong>Processing</strong> Challenges in <strong>Color</strong> <strong>Flow</strong> Imaging ....................................................... 18<br />
1.5.1 Background Considerations ........................................................................................ 18<br />
1.5.2 Challenges in Clutter Suppression............................................................................... 20<br />
Doppler Ensemble Size Limitations.................................................................................. 20<br />
Wideband Clutter Problems.............................................................................................. 20<br />
1.5.3 Challenges in Velocity Estimation............................................................................... 22<br />
- vi -
Table of Contents<br />
vii<br />
Carrier Frequency Variations .......................................................................................... 22<br />
Clutter Filter Distortions .................................................................................................. 23<br />
1.6 Concluding Remarks........................................................................................................... 24<br />
CHAPTER 2 Clutter Suppression in <strong>Color</strong> <strong>Flow</strong> Imaging: A Hankel-SVD Approach ...... 25<br />
2.1 Chapter Overview ............................................................................................................... 25<br />
2.2 Existing Clutter Filter Designs............................................................................................ 26<br />
2.2.1 Non-Adaptive Filters.................................................................................................... 26<br />
Background Considerations.............................................................................................. 26<br />
FIR and IIR Filters............................................................................................................ 27<br />
Regression Filters ............................................................................................................. 29<br />
2.2.2 Adaptive Filters............................................................................................................ 30<br />
Background Considerations.............................................................................................. 30<br />
Clutter-Downmixing Filters.............................................................................................. 31<br />
<strong>Eigen</strong>-<strong>Based</strong> Filters .......................................................................................................... 32<br />
2.2.3 Comparison of Clutter Filters...................................................................................... 33<br />
2.3 The Hankel-SVD Filter....................................................................................................... 35<br />
2.3.1 Background Considerations ........................................................................................ 35<br />
Design Motivations ........................................................................................................... 35<br />
Fundamental Principles.................................................................................................... 36<br />
2.3.2 Theoretical Formulation.............................................................................................. 36<br />
Construction of Hankel Data Matrix ................................................................................ 36<br />
SVD Analysis..................................................................................................................... 37<br />
Estimation of Clutter <strong>Eigen</strong>-Space Dimension ................................................................. 38<br />
Computational Considerations ......................................................................................... 40<br />
2.4 Simulation Method.............................................................................................................. 41<br />
2.4.1 <strong>Signal</strong> Synthesis Model ................................................................................................ 41<br />
General Principles............................................................................................................ 41<br />
Synthesis Procedure.......................................................................................................... 42<br />
Model Parameters............................................................................................................. 44<br />
2.4.2 Method of Analysis....................................................................................................... 46<br />
General Overview ............................................................................................................. 46<br />
Per<strong>for</strong>mance Measure....................................................................................................... 47<br />
2.5 Simulation Results .............................................................................................................. 47<br />
2.5.1 Characteristics of Principal Hankel Components ....................................................... 47<br />
Arterial <strong>Flow</strong> Scenario ..................................................................................................... 47<br />
Low-Velocity <strong>Flow</strong> Scenario............................................................................................. 49<br />
2.5.2 Post-Filter Clutter-to-Blood <strong>Signal</strong> Ratios.................................................................. 50<br />
Arterial <strong>Flow</strong> Scenario ..................................................................................................... 50<br />
Low-Velocity <strong>Flow</strong> Scenario............................................................................................. 51<br />
2.6 Concluding Remarks........................................................................................................... 53<br />
CHAPTER 3 Velocity Estimation in <strong>Color</strong> <strong>Flow</strong> Imaging: A Matrix Pencil Approach .... 54<br />
3.1 Chapter Overview ............................................................................................................... 54<br />
3.2 Existing <strong>Flow</strong> Estimation Strategies................................................................................... 55<br />
3.2.1 Non-Parametric Estimators......................................................................................... 55<br />
Fundamental Considerations............................................................................................ 55<br />
Frequency-<strong>Based</strong> <strong>Methods</strong>................................................................................................ 55
Table of Contents<br />
viii<br />
Time-Shift-<strong>Based</strong> <strong>Methods</strong>................................................................................................ 56<br />
3.2.2 Parametric Estimators................................................................................................. 57<br />
Fundamental Considerations............................................................................................ 57<br />
Autoregressive Modeling .................................................................................................. 58<br />
Multiple <strong>Signal</strong> Classification........................................................................................... 60<br />
3.3.2 Comparison of Frequency-<strong>Based</strong> Velocity Estimators................................................ 62<br />
3.3 The Matrix Pencil Estimator............................................................................................... 63<br />
3.3.1 Background Considerations ........................................................................................ 63<br />
Design Motivations ........................................................................................................... 63<br />
Fundamental Principles.................................................................................................... 64<br />
3.3.2 Theoretical Formulation.............................................................................................. 64<br />
Construction of Data Matrix Pair..................................................................................... 64<br />
Low-Rank Matrix Reduction............................................................................................. 67<br />
Numerical Computation.................................................................................................... 68<br />
Relationship to <strong>Eigen</strong>-Analysis......................................................................................... 70<br />
3.3.3 Applications in <strong>Color</strong> <strong>Flow</strong> Data <strong>Processing</strong>.............................................................. 71<br />
Rank-One Matrix Pencil ................................................................................................... 71<br />
Rank-Two Matrix Pencil................................................................................................... 72<br />
Rank-Adaptive Matrix Pencil............................................................................................ 73<br />
3.4 Simulation <strong>Methods</strong>............................................................................................................ 75<br />
3.4.1 Method of Study ........................................................................................................... 75<br />
General Overview ............................................................................................................. 75<br />
Per<strong>for</strong>mance Measures ..................................................................................................... 75<br />
3.4.2 Simulation Parameters................................................................................................. 76<br />
Data Synthesis................................................................................................................... 76<br />
Data Analysis.................................................................................................................... 77<br />
3.5 Simulation Results .............................................................................................................. 77<br />
3.5.1 <strong>Flow</strong> Scenario with Stationary Clutter ........................................................................ 77<br />
Per<strong>for</strong>mance of Low-Rank Reduction............................................................................... 77<br />
Comparative Assessment at Different BSNRs................................................................... 78<br />
Comparative Assessment at Different Blood Velocities.................................................... 81<br />
3.5.2 <strong>Flow</strong> Scenario with Non-Stationary Clutter ................................................................ 81<br />
Efficacy of Adaptive Rank Selection ................................................................................. 81<br />
Comparative Assessment at Different Blood Velocities.................................................... 83<br />
3.6 Concluding Remarks........................................................................................................... 84<br />
Single-Module Approach to <strong>Color</strong> <strong>Flow</strong> Data <strong>Processing</strong>:<br />
An Experimental Study ...................................................................................... 86<br />
4.1 Chapter Overview ............................................................................................................... 86<br />
4.2 The Single-Module <strong>Color</strong> <strong>Flow</strong> Data Processor................................................................. 87<br />
4.2.1 Background Considerations ........................................................................................ 87<br />
Design Motivation............................................................................................................. 87<br />
Fundamental Principles.................................................................................................... 88<br />
4.2.2 Processor Details......................................................................................................... 88<br />
Conceptual Overview........................................................................................................ 88<br />
Description of <strong>Processing</strong> Module.................................................................................... 89<br />
4.3 In Vitro <strong>Color</strong> <strong>Flow</strong> Imaging Study.................................................................................... 91<br />
CHAPTER 4
Table of Contents<br />
ix<br />
4.3.1 Experimental Protocol................................................................................................. 91<br />
Data Acquisition ............................................................................................................... 91<br />
<strong>Signal</strong> <strong>Processing</strong> Procedure............................................................................................ 92<br />
Data Analysis Method....................................................................................................... 93<br />
4.3.2 Experimental Results ................................................................................................... 93<br />
<strong>Color</strong> <strong>Flow</strong> Images ........................................................................................................... 93<br />
Velocity Profiles................................................................................................................ 94<br />
4.4 In Vivo <strong>Color</strong> <strong>Flow</strong> Imaging Studies.................................................................................. 96<br />
4.4.1 Experimental Protocol................................................................................................. 96<br />
Data Acquisition ............................................................................................................... 96<br />
<strong>Signal</strong> <strong>Processing</strong> Procedure............................................................................................ 96<br />
Data Analysis Method....................................................................................................... 97<br />
4.4.2 Case Study with Vessel Wall Motion ........................................................................... 97<br />
Overview of Scenario........................................................................................................ 97<br />
Results: <strong>Color</strong> <strong>Flow</strong> Images.............................................................................................. 98<br />
Results: Regional Power Comparison............................................................................ 100<br />
4.4.3 Case Study with Translational Tissue Motion ........................................................... 100<br />
Overview of Scenario...................................................................................................... 100<br />
Results: <strong>Color</strong> <strong>Flow</strong> Images............................................................................................ 102<br />
Results: Regional Power Comparison............................................................................ 104<br />
4.5 Concluding Remarks......................................................................................................... 105<br />
CHAPTER 5 Thesis Summary and Future Directions......................................................... 106<br />
5.1 Chapter Overview ............................................................................................................. 106<br />
5.2 Summary of Thesis Study................................................................................................. 106<br />
5.2.1 Overall Scope............................................................................................................. 106<br />
5.2.2 Specific Developments ............................................................................................... 107<br />
Hankel-SVD Filter .......................................................................................................... 107<br />
Matrix Pencil .................................................................................................................. 108<br />
Single-Module <strong>Color</strong> <strong>Flow</strong> Data Processor ................................................................... 109<br />
Doppler <strong>Signal</strong> Synthesis Model..................................................................................... 110<br />
5.2.3 List of Primary Research Contributions.................................................................... 111<br />
5.3 Future Research Directions............................................................................................... 111<br />
5.3.1 Developmental Aspects .............................................................................................. 111<br />
Real-Time Computation Issues ....................................................................................... 111<br />
Adaptive Thresholding.................................................................................................... 112<br />
5.3.2 Experimental Aspects................................................................................................. 112<br />
<strong>Flow</strong> Imaging Phantom Design ...................................................................................... 112<br />
Doppler Clutter Analysis ................................................................................................ 113<br />
5.3.3 Theoretical Aspects.................................................................................................... 113<br />
<strong>Processing</strong> of Non-Uni<strong>for</strong>mly Sampled <strong>Color</strong> <strong>Flow</strong> Data.............................................. 113<br />
Contrast-Enhanced <strong>Color</strong> <strong>Flow</strong> Data <strong>Processing</strong>.......................................................... 114<br />
Other Single-Snapshot-<strong>Based</strong> <strong>Flow</strong> Estimators.............................................................. 114<br />
<strong>Flow</strong> Analysis Using Image <strong>Processing</strong> Tools................................................................ 115<br />
5.4 Final Remarks ................................................................................................................... 115<br />
REFERENCES ............................................................................................................................. 117
LIST OF ABBREVIATIONS AND SYMBOLS<br />
Abbreviations<br />
2D Two-dimensional<br />
3D Three-dimensional<br />
ACR Autocorrelator<br />
AR Autoregressive<br />
B-mode Brightness-mode<br />
BSNR Blood-signal-to-noise ratio<br />
BV Blood vessel<br />
CBR Clutter-to-blood signal ratio<br />
CW Continuous-wave<br />
ESPRIT Estimation of signal parameters via rotational invariance techniques<br />
FIR Finite impulse response<br />
GE Generalized eigenvalue<br />
IIR Infinite impulse response<br />
I/Q In-phase/quadrature<br />
KL Karhunen-Loeve<br />
M-mode Motion-mode<br />
MP Matrix Pencil<br />
MT Moving tissue<br />
MUSIC Multiple signal classification<br />
PW Pulsed-wave<br />
RK-1 MP Rank-one Matrix Pencil<br />
RK-2 MP Rank-two Matrix Pencil<br />
RK-A MP Rank-adaptive Matrix Pencil<br />
RMS Root-mean-squared<br />
ST Static tissue<br />
SVD Singular value decomposition<br />
- x -
List of Symbols and Abbreviations<br />
xi<br />
Physical and Imaging Parameters<br />
α dB <strong>Ultrasound</strong> attenuation factor (in dB scale)<br />
Δφ Inter-pulse phase change<br />
Δt ( =T PRI ) Inter-pulse period (essentially the pulse repetition interval)<br />
λ o ( = c/f ) Wavelength of ultrasound<br />
θ Beam-flow angle<br />
c o <strong>Ultrasound</strong> propagation speed<br />
d max Maximum imaging depth of color flow image<br />
d pen Penetration depth of ultrasound<br />
d res Depth (or axial) resolution of color flow image<br />
f D Doppler shift (frequency)<br />
f D(res) Doppler spectral resolution<br />
f o Transmitted ultrasound frequency<br />
f r Received ultrasound frequency<br />
F Focal length<br />
I pen(dB) Intensity attenuation (in dB) at a given penetration depth<br />
l res Beam line (or lateral) resolution of color flow image<br />
N cyc Number of cycles in transmit pulse<br />
N D Doppler ensemble size (or number of firings per beam line)<br />
N L Number of beam lines in color flow image<br />
t D Inter-pulse arrival time shift<br />
T F Frame period of color flow image<br />
T P Temporal length of transmit pulse<br />
T PRI Pulse repetition interval<br />
v<br />
Scatterer velocity<br />
v alias Velocity aliasing limit of color flow image<br />
v res Velocity resolution of color flow image<br />
W Aperture width<br />
Doppler <strong>Signal</strong> <strong>Processing</strong> Parameters<br />
χ k<br />
Fitting coefficient of the k th basis vector in a regression filter or principalcomponent<br />
signal model<br />
Kronecker delta function that equals to one when n = i+j<br />
δ n=i+j
List of Symbols and Abbreviations<br />
xii<br />
δ n=j–i<br />
Δf thr<br />
γ k<br />
λ<br />
λ k<br />
ρ y<br />
σ k<br />
φ(n)<br />
ϕ k (n)<br />
a(i,j)<br />
a k (i,j)<br />
b k<br />
c k<br />
e n (i,j)<br />
f D,k<br />
f D,kb<br />
f D,thr<br />
f D(c)<br />
f D(est)<br />
f thr(clut)<br />
H(f D )<br />
K<br />
K c<br />
K nom<br />
M<br />
P<br />
R y (0)<br />
R y (1)<br />
Kronecker delta function that equals to one when n = j–i<br />
Clutter bandwidth threshold in the Hankel-SVD filter; also the spectral<br />
spread threshold in rank-adaptive Matrix Pencil<br />
Coefficient of the k th basis vector in a KL expansion or Hankel component<br />
approximation<br />
<strong>Eigen</strong>value<br />
<strong>Eigen</strong>value (expected value) of the k th basis vector in a KL expansion<br />
Post-filter Doppler signal power<br />
k th singular value in a SVD<br />
Instantaneous phase of carrier in a clutter-downmixing filter<br />
n th sample of the k th basis vector in a Hankel-SVD<br />
(i,j) th entry in the Hankel data matrix<br />
(i,j) th entry in the k th rank-one Hankel component<br />
Weight of the k th past input sample in a FIR/IIR filter<br />
Weight of the k th past output sample in an IIR filter; coefficient of an AR or<br />
MUSIC characteristic polynomial<br />
(i,j) th H<br />
entry of the noise subspace projection matrix E n E n<br />
Mean Doppler frequency of k th principal basis in a Hankel-SVD or principalcomponent<br />
signal model<br />
Principal frequency corresponding to the modal flow component<br />
Doppler frequency threshold in the Hankel-SVD filter<br />
Mean Doppler frequency of clutter (used <strong>for</strong> clutter downmixing)<br />
Doppler frequency estimate (mean or mode)<br />
Spectral threshold <strong>for</strong> checking clutter presence (used in Hankel-SVD filter)<br />
Filter response at a given Doppler frequency f D<br />
Order <strong>for</strong> a filter or eigen-structure rank <strong>for</strong> a principal-component signal<br />
model<br />
Estimated clutter eigen-space dimension in the Hankel-SVD filter<br />
Nominal eigen-structure rank used by rank-adaptive Matrix Pencil<br />
Number of snapshots used to estimate Doppler correlation matrix<br />
Dimension parameter of a Hankel data matrix or matrix pencil<br />
Autocorrelation function at zero lag<br />
Autocorrelation function at single sample lag
List of Symbols and Abbreviations<br />
xiii<br />
w(n)<br />
x(n)<br />
x DM (n)<br />
X AR (f D )<br />
X MUSIC (f D )<br />
y(n)<br />
z k<br />
Δ<br />
ϕ k<br />
Θ<br />
Φ<br />
a 0<br />
a 1<br />
A<br />
A 0<br />
A 1<br />
Ã<br />
à 0<br />
à 1<br />
A k<br />
b<br />
c<br />
ĉ<br />
C<br />
D<br />
e k<br />
E n<br />
F<br />
I<br />
p k<br />
q<br />
R x<br />
n th sample of white noise in an ensemble<br />
n th sample of Doppler signal in an ensemble<br />
n th sample of the downmixed Doppler signal<br />
<strong>Signal</strong> spectrum <strong>for</strong> an AR model at a given Doppler frequency f D<br />
<strong>Signal</strong> spectrum <strong>for</strong> a MUSIC model at a given Doppler frequency f D<br />
n th sample of filtered Doppler signal in an ensemble<br />
k th characteristic mode in a principal-component signal model<br />
Noise perturbation matrix in a low-rank matrix reduction<br />
k th basis vector in a Hankel component approximation<br />
Diagonal matrix operator <strong>for</strong> clutter downmixing<br />
Diagonal matrix of GEs in the decomposition proof of Matrix Pencil<br />
Data vector <strong>for</strong> first N D –1 samples in an ensemble<br />
Data vector <strong>for</strong> last N D –1 samples in an ensemble<br />
Hankel data matrix<br />
Hankel data matrix <strong>for</strong> the first N D –1 samples in an ensemble<br />
Hankel data matrix <strong>for</strong> the last N D –1 samples in an ensemble<br />
Rank-reduced Hankel data matrix in a low-rank approximation<br />
First N D –1 rows of a rank-reduced Hankel data matrix<br />
Last N D –1 rows of a rank-reduced Hankel data matrix<br />
k th rank-one Hankel component in a SVD<br />
Blood component of Doppler signal vector<br />
Clutter component of Doppler signal vector<br />
Clutter model vector of a regression filter<br />
Clutter space matrix of a regression filter<br />
Diagonal matrix of model coefficients in the decomposition of Matrix Pencil<br />
k th basis vector in a KL expansion<br />
Noise subspace matrix in a KL expansion<br />
Filter matrix<br />
Identity matrix<br />
k th -order polynomial basis vector in a regression filter<br />
<strong>Eigen</strong>vector<br />
Doppler correlation matrix
List of Symbols and Abbreviations<br />
xiv<br />
S Steady-state filter matrix <strong>for</strong> the state-space <strong>for</strong>m of FIR/IIR filter<br />
T Transient filter matrix <strong>for</strong> the state-space <strong>for</strong>m of FIR/IIR filter<br />
u k k th left singular vector in a SVD<br />
u 0,k First N D –P samples of the k th left singular vector in a Hankel-SVD<br />
u 1,k Last N D –P samples of the k th left singular vector in a Hankel-SVD<br />
v k k th right singular vector in a SVD<br />
v CS Complex sinusoid vector<br />
v CS(k)<br />
Complex sinusoid vector <strong>for</strong> the k th component in a principal-component<br />
signal model<br />
v init Initial state vector <strong>for</strong> the state-space <strong>for</strong>m of FIR/IIR filter<br />
w White noise component of Doppler signal vector<br />
w f White noise component in a filtered Doppler signal vector<br />
x<br />
Doppler signal vector<br />
x m m th Doppler signal vector used to estimate Doppler correlation matrix<br />
x DM Doppler signal vector after clutter downmixing<br />
y<br />
Filtered Doppler signal vector<br />
Z L Left complex-sinusoid matrix in the decomposition proof of Matrix Pencil<br />
Z R Right complex-sinusoid matrix in the decomposition proof of Matrix Pencil<br />
Doppler <strong>Signal</strong> Modeling Parameters<br />
β b Bandwidth of the blood signal template<br />
φ c,max Peak clutter phase deviation in the Doppler signal synthesis model<br />
κ b Blood magnitude coefficient in the synthesized signal<br />
κ c Clutter magnitude coefficient in the synthesized signal<br />
κ w Noise magnitude coefficient in the synthesized signal<br />
f D(b) Modulating frequency of the blood signal template<br />
f vib Clutter vibration frequency<br />
g b (n) n th random sample in the blood scatterer distribution function<br />
g c Random sample used to model clutter scattering distribution<br />
h b (n) n th sample in the blood signal template of signal synthesis model<br />
h c (n) n th sample in the clutter signal template of signal synthesis model<br />
Temporal width of the blood signal template<br />
T dur
CHAPTER 1<br />
Introduction to <strong>Color</strong> <strong>Flow</strong> Imaging<br />
1.1 Clinical Background<br />
Cardiovascular diseases are health disorders related to the functioning of the heart and the<br />
blood vessels. According to statistics from the Heart and Stroke Foundation of Canada (2003),<br />
this class of health problems is the leading cause of death in Canada, affecting over 70,000<br />
Canadians every year and costing the Canadian economy $18.4 billion annually due to direct<br />
hospitalization charges and lost productivity. In terms of its causes, vascular stenoses and<br />
strokes have contributed to 84% of cardiovascular patient cases. As such, detailed diagnoses are<br />
often per<strong>for</strong>med to study the cardiovascular flow dynamics inside a patient. From the results<br />
acquired through these diagnoses, clinicians can then deploy suitable therapies and surgeries to<br />
the patient as necessary.<br />
Imaging studies of cardiovascular flow dynamics are traditionally per<strong>for</strong>med using X-<br />
rays. In fact, Nederkoorn et al. (2003) pointed out in their recent review that X-ray angiography,<br />
which involves the insertion of a catheter device into the patient’s vasculature, still remains as<br />
the gold standard <strong>for</strong> such diagnosis. However, as an invasive modality, this imaging approach<br />
carries the risk of accidentally dislodging fatty tissues and calcified deposits from vessel walls.<br />
Over the past twenty years, new non-invasive imaging tools such as ultrasound color flow<br />
imaging and magnetic resonance angiography have been developed in attempt to obtain the same<br />
diagnostic in<strong>for</strong>mation without intrusion to the patient’s vasculature. As suggested by Beach et<br />
al. (1997), these two new imaging tools have similar spatial imaging resolution and appear to be<br />
complementary methods of each other. Nevertheless, ultrasound color flow imaging has better<br />
real-time scanning capabilities and is more af<strong>for</strong>dable by most clinics.<br />
Ever since its development, ultrasound color flow imaging has been used <strong>for</strong> various<br />
types of cardiovascular flow studies. As reviewed by Ferrara and DeAngelis (1997), it has been<br />
used to identify the presence of vascular stenoses, assess complicated flow patterns, diagnose <strong>for</strong><br />
the development of aneurysms and tumors, study the patency of implanted shunts, and visualize<br />
- 1 -
Chapter 1. Introduction to <strong>Ultrasound</strong> <strong>Color</strong> <strong>Flow</strong> Imaging 2<br />
blood regurgitations between heart chambers during a cardiac cycle. More recently, this imaging<br />
tool has found new applications such as microcirculation assessment in the eyes (Williamson and<br />
Harris 1996), treatment response monitoring during cancer therapies (Lagalla et al. 1998),<br />
contractility examination of heart muscles (Waggoner and Bierig 2001), and guidance of<br />
interventional devices (Armstrong et al. 2001). From these clinical applications, it can be seen<br />
that ultrasound color flow imaging is often used as a high-level vascular visualization tool.<br />
1.2 Outline of Thesis Study<br />
1.2.1 Research Motivation and Hypothesis<br />
Given the widespread use of ultrasound color flow imaging in medical diagnostics, it is<br />
important <strong>for</strong> these images to provide accurate visualization of the vascular flow dynamics. As<br />
suggested in some evaluation studies (e.g. Arbeille et al. 1999, Stewart 2001), failure to provide<br />
accurate flow in<strong>for</strong>mation in color flow images may lead to an increased risk of misdiagnosis<br />
and may cause assessment difficulties during long-term patient monitoring. However, from a<br />
signal processing perspective, there are several factors that tend to reduce the accuracy of blood<br />
flow estimates displayed in color flow images. As will be described in Section 1.5, these sources<br />
of errors can be divided into two categories: 1) those that arise during suppression of clutter in<br />
the color flow data, and 2) those that arise during estimation of flow velocities. In terms of their<br />
impact, these error sources may lead to the appearance of spurious or erroneous pixels in the<br />
color flow images. In turn, they tend to obscure visualization of the actual flow dynamics.<br />
To address the challenges in color flow signal processing, it is necessary to develop<br />
suitable methods that can adequately suppress clutter and derive flow estimates with minimal<br />
bias. We hypothesize that a more effective clutter suppression method can be developed through<br />
the use of eigen-based approaches like principal component analysis, while a more accurate flow<br />
estimation strategy can be <strong>for</strong>mulated through parametric spectral analysis along with adaptive<br />
rank selection. With the advent of improved clutter filters and flow estimators, we can anticipate<br />
that the clinical reliability of color flow images will be improved.<br />
1.2.2 Goals and Objectives<br />
Driven by the research hypothesis, the overall goal of this thesis study is to develop new<br />
eigen-based signal processing methods <strong>for</strong> color flow imaging applications. Specifically, it is
Chapter 1. Introduction to <strong>Ultrasound</strong> <strong>Color</strong> <strong>Flow</strong> Imaging 3<br />
our intent to devise a new clutter filter design and a novel flow estimation strategy by exploiting<br />
the eigen-space properties of two matrix <strong>for</strong>ms known as the Hankel matrix and the matrix<br />
pencil. We will also consider how these two eigen-based methods can be combined to create a<br />
new color flow data processor that can per<strong>for</strong>m flow detection and flow estimation in parallel<br />
within the same processing module.<br />
To achieve our overall research goal, this thesis study has been broken down into the<br />
following three specific objectives:<br />
• Design a new eigen-based clutter filter (referred to as Hankel-SVD filter) based on Hankel<br />
matrices created from the acquired signals and evaluate its efficacy in clutter suppression;<br />
• Develop a new eigen-based flow estimator (referred to as Matrix Pencil estimator) from the<br />
properties of matrix pencil and investigate its theoretical estimation per<strong>for</strong>mance;<br />
• Integrate the above two designs together to <strong>for</strong>m a new color flow signal processor and study<br />
its per<strong>for</strong>mance under various in vitro and in vivo flow scenarios.<br />
Note that the signal processing strategies listed in the first two objectives can be considered as<br />
the building blocks <strong>for</strong> the new signal processor mentioned in the last objective.<br />
1.2.3 Chapter Overview<br />
The remainder of this introductory chapter is intended to provide an overview on the<br />
principles of color flow imaging and describe the signal processing problems being addressed in<br />
this thesis. In particular, the chapter contents have been organized as follows:<br />
• Section 1.3 provides a general overview on how ultrasound can be used to study blood flow<br />
dynamics and describes the display modes in ultrasound-based flow estimation methods;<br />
• Section 1.4 gives a system-level description of ultrasound color flow imaging, discusses the<br />
various imaging parameters that need to be considered, and surveys some advanced color<br />
flow imaging schemes reported in the literature;<br />
• Section 1.5 discusses the major signal processing challenges in ultrasound color flow<br />
imaging and reviews existing ways on how these challenges are being addressed;<br />
• Section 1.6 summarizes the research problem and outlines the organization of the subsequent<br />
chapters in this thesis.<br />
Readers who are familiar with color flow imaging principles are encouraged to proceed<br />
directly to Section 1.5 which describes the sources of errors in color flow signal processing.
Chapter 1. Introduction to <strong>Ultrasound</strong> <strong>Color</strong> <strong>Flow</strong> Imaging 4<br />
1.3 Overview of Doppler <strong>Ultrasound</strong><br />
1.3.1 General Concepts<br />
Physical Principles<br />
As recently reviewed by Cobbold (2006, Sec. 9.1 and 10.1), flow velocity estimation<br />
using ultrasound was first successfully attempted in the late 1950s. This approach involves the<br />
transmission of ultrasound waves (i.e. acoustic waves with frequencies beyond the human<br />
audible range of 20 kHz) into the blood vessel and the subsequent reception of echoes returned<br />
from red blood cell scatterers flowing inside the vessel. In general, there are two excitation<br />
modes in ultrasound-based flow estimation methods: the continuous-wave (CW) mode and the<br />
pulsed-wave (PW) mode. As revealed by its name, the CW mode is based on the use of<br />
continuous ultrasound waves, and it estimates flow velocities by measuring the Doppler<br />
frequency shifts (i.e. changes in the perceived frequency due to source or target motion) between<br />
the transmitted wave and the received echoes. Since the Doppler Effect is exploited, this<br />
excitation mode is often referred to as CW Doppler ultrasound (or simply called CW Doppler).<br />
On the other hand, the PW mode is based on the use of finite-duration ultrasound pulses, and it<br />
estimates flow velocities by measuring the time shifts between pulse echoes. This excitation<br />
mode is historically referred to as PW Doppler ultrasound (or simply called PW Doppler), even<br />
though it is somewhat of a misnomer † .<br />
Doppler Equation<br />
For both CW and PW Doppler ultrasound, the scatterer velocity distribution is one of the<br />
factors that the shift magnitudes depend on (regardless of whether they are Doppler shifts or time<br />
shifts). Since vascular flow dynamics in turn govern the scatterer velocity distribution, it is<br />
possible to gain knowledge about the flow profile from studying the induced shift magnitudes.<br />
This principle essentially <strong>for</strong>ms the working basis of Doppler ultrasound. In terms of the other<br />
factors that govern the shift magnitudes, the Doppler frequency shifts seen in the backscattered<br />
echoes of CW Doppler also depend on the beam-flow angle θ, transmitted ultrasound frequency<br />
f o , and acoustic propagation speed c o . The equation that relates the Doppler shift f D to the various<br />
† In early research studies, it was believed that the arrival time shifts are induced by the classical Doppler Effect.<br />
Hence, the “Doppler” misnomer has evolved. Further discussion on this misnomer can be found in many Doppler<br />
ultrasound textbooks (e.g. see Jensen 1996, Ch. 4; Evans and McDicken 2000, Ch. 4; Cobbold 2006, Ch. 10).
Chapter 1. Introduction to <strong>Ultrasound</strong> <strong>Color</strong> <strong>Flow</strong> Imaging 5<br />
contributing factors is often referred to as the Doppler equation, and it has the following<br />
mathematical <strong>for</strong>m (Cobbold 2006, Sec. 9.1):<br />
f<br />
D<br />
2vf<br />
o<br />
= f<br />
o<br />
− f<br />
r<br />
≈ cosθ (<strong>for</strong> c o >> v), (1-1)<br />
c<br />
o<br />
where v is the scatterer velocity and f r is the received ultrasound frequency.<br />
In comparison to CW Doppler, the time shifts in PW Doppler ultrasound are also<br />
dependent on similar factors even though the two excitation modes are based on different<br />
physical principles. As derived in various ultrasound textbooks (e.g. see Cobbold 2006, Sec.<br />
10.2), the time shift t D between pulse echoes can be described by the following expression:<br />
2vTPRI<br />
t<br />
D<br />
= cosθ , (1-2)<br />
c<br />
o<br />
where T PRI is the pulse repetition interval. Note that the time shifts seen in PW Doppler can also<br />
be interpreted as a frequency-shift mechanism. In particular, given that frequency is equal to the<br />
phase change over time by definition, the time shift between pulse echoes carries the following<br />
frequency interpretation:<br />
Δφ<br />
f<br />
ot<br />
≈<br />
Δt<br />
T<br />
D<br />
PRI<br />
2vf<br />
=<br />
c<br />
o<br />
o<br />
cosθ<br />
, (1-3)<br />
where Δφ = f o t D indicates the phase change between pulse echoes. Comparing between (1-1) and<br />
(1-3), it can be seen that the frequency shift in PW Doppler ultrasound is essentially equivalent<br />
to the classical Doppler shift observed in CW Doppler. Hence, without loss of generality, we<br />
will interchangeably use the symbol f D and the term “Doppler frequency” throughout this thesis<br />
to denote the frequency shifts in PW Doppler ultrasound as well.<br />
Range Localization<br />
Even though it is more complicated, PW Doppler ultrasound is often more preferred<br />
because of its range localization capabilities. In particular, by using a range gate to separate the<br />
echoes originating from outside scatterers, this excitation mode can provide analysis of the flow<br />
velocity distribution at specific ranges of interest (Cobbold 2006, Sec. 10.3). Such range<br />
localization is useful to blood flow studies since it facilitates the identification of local changes<br />
in the flow profile inside the vasculature. Hence, much ef<strong>for</strong>t has been devoted to the<br />
development of PW Doppler ultrasound over the years, but it should be noted that CW Doppler
Chapter 1. Introduction to <strong>Ultrasound</strong> <strong>Color</strong> <strong>Flow</strong> Imaging 6<br />
(b) Velocity Map<br />
(a) Spectrogram<br />
(c) <strong>Flow</strong> Power Map<br />
Fig. 1-1. Various display modes in Doppler ultrasound: (a) spectrogram (time-varying<br />
velocity distribution from one sample volume); (b) velocity map (directional flow estimates<br />
within an image plane in red and blue hues); (c) flow power map (blood signal power within<br />
an image plane in pale-red hues). These figures were reproduced with permission from<br />
ZONARE Medical Systems (Mountain View, CA, USA).<br />
ultrasound still remains of use in studies where range localization is not important. In this thesis,<br />
we shall focus the attention on PW Doppler ultrasound, and from hereon, it will be implicitly<br />
assumed that PW excitations are used.<br />
1.3.2 Display Modes<br />
Spectrogram<br />
One commonly used display mode in Doppler ultrasound is a spectrogram (or spectral<br />
Doppler display) that shows the time-varying flow velocity distribution derived from the<br />
measured Doppler frequencies. An example of this type of display is shown in Fig. 1-1a. Note<br />
that the presented flow in<strong>for</strong>mation is derived from a single sample volume whose size can be<br />
readily adjusted by the user during a diagnostic procedure. Also, the placement of the sample
Chapter 1. Introduction to <strong>Ultrasound</strong> <strong>Color</strong> <strong>Flow</strong> Imaging 7<br />
volume is guided by a grayscale ultrasound image (i.e. a brightness-mode or B-mode image) that<br />
reveals the anatomical structure within the ultrasound transducer’s field of view. In terms of the<br />
spectrogram’s significance, clinicians often use this display mode to gain quantitative insights on<br />
the vascular flow dynamics within a certain range of interest inside the blood vessel. In<br />
particular, since vascular stenoses often give rise to disturbed and even turbulent flow conditions,<br />
clinicians can study the spectrogram to assess the healthiness of a patient’s vasculature.<br />
<strong>Color</strong> <strong>Flow</strong> Image<br />
Another mode of display <strong>for</strong> Doppler ultrasound is a color flow image (or velocity map)<br />
that shows flow velocity estimates of different sample volumes within the transducer’s field of<br />
view. An example of this type of display is illustrated in Fig. 1-1b. As shown in the figure, the<br />
velocity estimates are displayed in the <strong>for</strong>m of color pixels, and they are superimposed onto a B-<br />
mode image of the anatomical structure. This duplex image display <strong>for</strong>mat allows clinicians to<br />
easily identify blood flow within an entire imaging view. In comparison to the spectrogram that<br />
shows the flow distribution, the color flow image only provides a single velocity estimate<br />
(usually the mean or modal velocity) <strong>for</strong> each sample volume. As such, during a diagnostic<br />
procedure, the color flow image is usually used as a high-level flow visualization tool prior to<br />
detailed diagnoses at specific parts of the vasculature.<br />
For the color flow image, it is possible to display flow power estimates instead of flow<br />
velocities. The resulting duplex images are often referred to as flow power maps (or Doppler<br />
power maps), and an example of this type of color flow image is given in Fig. 1-1c. Note that, at<br />
large beam-flow angles, the flow power map has diagnostic advantage over the velocity map<br />
because the echo power returned from blood scatterers is less dependent of the beam-flow angle<br />
(while the Doppler shift magnitude is angle-dependent). On the other hand, the velocity map has<br />
the advantage of being able to provide directional flow in<strong>for</strong>mation that is clinically important<br />
<strong>for</strong> the identification of arteries and veins.<br />
M-Mode Profile<br />
A third but less common mode of display <strong>for</strong> Doppler ultrasound is a motion-mode (or<br />
M-mode) profile that presents color-coded flow in<strong>for</strong>mation over time along one ultrasound<br />
beam line. In connection with the spectrogram and the color flow image, the M-mode profile<br />
can be considered as a hybrid display mode because it gives time-varying velocity estimates like
Chapter 1. Introduction to <strong>Ultrasound</strong> <strong>Color</strong> <strong>Flow</strong> Imaging 8<br />
Wave<strong>for</strong>m<br />
Oscillator<br />
Pulser<br />
Echo Amplifier<br />
Mixer<br />
Transducer<br />
Lowpass Filter<br />
Surface<br />
Doppler<br />
lateral<br />
(or time)<br />
Image View<br />
Artery<br />
depth<br />
Sampling Memory<br />
(depth res. d res )<br />
(Doppler res. T PRI )<br />
(image mode: lateral res. l res )<br />
(M-mode: time res. N D T PRI )<br />
To <strong>Flow</strong><br />
Estimation<br />
Fig. 1-2. Overview of the data acquisition setup used <strong>for</strong> ultrasound color flow imaging.<br />
Note that the sampling memory in the image mode has three dimensions: depth, lateral, and<br />
Doppler. In M-mode, the lateral dimension becomes the time dimension since the multipulse<br />
data acquisition process is repeated over time along a single beam line.<br />
a spectrogram whilst providing multi-depth flow in<strong>for</strong>mation like a color flow image. The<br />
advantage of using this display mode is that it allows clinicians to study how the flow profile<br />
temporally evolves through a particular cross-section of a blood vessel. Note that the velocity<br />
estimates provided in an M-mode profile are derived using the same methods as the ones used to<br />
produce color flow images. Also, instead of showing velocity estimates, it is possible to display<br />
flow power estimates in an M-mode profile.<br />
1.4 Principles of <strong>Ultrasound</strong> <strong>Color</strong> <strong>Flow</strong> Imaging<br />
1.4.1 System-Level Overview<br />
Data Acquisition<br />
From a system-level perspective, the <strong>for</strong>mation of each color flow image frame or M-<br />
mode profile can be divided into two main stages: data acquisition and signal processing. As<br />
illustrated in Fig. 1-2, the data acquisition stage first involves the use of an ultrasound transducer<br />
(often an array transducer) to transmit pulse firings and collect the returned echoes. In order to
Chapter 1. Introduction to <strong>Ultrasound</strong> <strong>Color</strong> <strong>Flow</strong> Imaging 9<br />
produce an image, the pulse firings are repeated <strong>for</strong> different beam lines within the field of view;<br />
in contrast, if an M-mode profile is to be <strong>for</strong>med, then pulses are only fired along a single beam<br />
line. Subsequently, the received echoes from the different pulse firings are individually<br />
downmixed to baseband by mixing the echoes with the carrier frequency of the transmitted pulse<br />
and applying a lowpass filter to retain only the baseband spectral components in the mixed<br />
echoes. For this downmixing process, the in-phase/quadrature (I/Q) demodulation scheme is<br />
often implemented so that the analytic <strong>for</strong>m of the pulse echoes can be obtained <strong>for</strong> subsequent<br />
analysis. After the pulse echoes are downmixed, they are sampled at time points that correspond<br />
to various depths in order to study the inter-pulse phase changes or time shifts at various axial<br />
and lateral positions (i.e. sample volumes) within the field of view. The resulting data samples<br />
can then be arranged into a three-dimensional (3D) data array: one dimension <strong>for</strong> pulse number,<br />
one <strong>for</strong> depth/axial position, and the other <strong>for</strong> line/lateral position. Since the pulse-number<br />
dimension of the 3D data array essentially contains in<strong>for</strong>mation on the inter-pulse phase changes,<br />
this dimension is also referred to as the Doppler axis † . As well, the ensemble of pulse echo<br />
samples <strong>for</strong> a particular sample volume is referred to as the Doppler signal.<br />
<strong>Signal</strong> <strong>Processing</strong><br />
In the signal processing stage of ultrasound color flow imaging, a Doppler power map<br />
and a flow velocity map are computed from the array of Doppler signals acquired through the<br />
pulse firings. Most commercial systems compute these two <strong>for</strong>ms of flow in<strong>for</strong>mation through a<br />
frequency-based estimation approach (i.e. by studying the inter-pulse phase changes) because of<br />
its relatively simpler computation complexity as compared to time-shift-based estimation<br />
approaches (Evans and McDicken 2000, Sec. 4.5.3). As shown in Fig. 1-3, the frequencydomain<br />
flow estimation process first involves the use of a digital highpass filter to suppress lowfrequency<br />
echoes that may be present in each Doppler signal. This filtering operation is carried<br />
out in attempt to distinguish the Doppler frequency shifts of blood scatterers from those that may<br />
arise due to acoustic reverberations of nearby tissues (whose strength can reach 60-80 dB greater<br />
than blood depending on the scanner’s dynamic range). After the filtered data samples are<br />
obtained, the average power of each post-filter Doppler signal is estimated by simply finding the<br />
† Some authors refer to the pulse-number dimension as the “slow-time” axis to avoid the Doppler misnomer. Also,<br />
the depth dimension is sometimes called the “fast-time” axis since it is the relative time axis of each pulse echo.
Chapter 1. Introduction to <strong>Ultrasound</strong> <strong>Color</strong> <strong>Flow</strong> Imaging 10<br />
Power Estimator<br />
lateral<br />
(or time)<br />
depth<br />
Power Map<br />
Memory<br />
Doppler<br />
<strong>Signal</strong><br />
Highpass Clutter<br />
Filter<br />
(cutoff freq. f D,thr )<br />
Spurious Pixel<br />
Removal<br />
Velocity<br />
Estimator<br />
lateral<br />
(or time)<br />
depth<br />
Velocity Map<br />
Memory<br />
Display<br />
Fig. 1-3. The signal processing scheme used in the frequency-domain flow estimation<br />
approach. Note that the post-filter Doppler power estimates are used <strong>for</strong> both generating the<br />
power map as well as deciding whether a map pixel corresponds to real flow in<strong>for</strong>mation.<br />
mean-squared signal value. The velocity estimate of each post-filter Doppler signal is also<br />
determined by first finding the mean Doppler frequency of the filtered signal and then converting<br />
this estimate into a velocity value via the Doppler equation given in (1-3) † . Once the power and<br />
velocity estimates are computed <strong>for</strong> all the Doppler data, they are turned into image pixels by<br />
respectively mapping the estimates onto a power map color scale and a velocity map color scale.<br />
Since blood flow may not be present in all places within the field of view, spurious pixels are<br />
removed from the power and velocity maps if their corresponding filtered Doppler power is<br />
below a threshold value. Finally, depending on which display mode is used, the resulting<br />
Doppler power map or flow velocity map is superimposed onto a B-mode image of the<br />
underlying structure to <strong>for</strong>m a duplex display.<br />
1.4.2 Imaging Considerations<br />
Transducer Type and Transmit Pulse Characteristics<br />
In color flow imaging, the ability to provide real-time visualization of flow dynamics is<br />
often considered as an important diagnostic feature. To realize this real-time functionality, it is<br />
necessary to use an ultrasound transducer that is capable of transmitting all the pulse firings and<br />
† During the conversion, the parameters θ, f o , and c o in the Doppler equation are assumed to be constant.
Chapter 1. Introduction to <strong>Ultrasound</strong> <strong>Color</strong> <strong>Flow</strong> Imaging 11<br />
collecting the returned echoes <strong>for</strong> each image frame within a short frame period (at most on the<br />
order of 100 ms). Hence, as similar to grayscale ultrasound imaging (i.e. B-mode scans), color<br />
flow imaging typically uses an array transducer such as linear array or curved array <strong>for</strong> the data<br />
acquisition. The advantage of using the array transducer is that it can acquire pulse echoes along<br />
various beam lines in real-time by sequentially exciting different groups of transducer elements.<br />
The array transducer is also able to acquire data at more acute beam-flow angles through<br />
electronic beam steering and in turn avoid Doppler frequency detection problems at neartransverse<br />
beam-flow angles. A further advantage of the array transducer is that it can acquire<br />
the pulse echoes from the different beam lines in an interleaved order to increase the coherence<br />
in the flow estimates between beam lines. In terms of their operating frequencies, commercially<br />
available array transducers generally operate in the frequency range between 1-17 MHz † . As<br />
will be discussed later in this section, such frequency range is generally sufficient <strong>for</strong> flow<br />
studies in large blood vessels, but higher frequencies are needed in order to study blood flow in<br />
the microcirculation. For data acquisition at higher ultrasound frequencies, current systems use a<br />
single-element transducer to obtain pulse echoes along one beam line and mechanically move the<br />
transducer to collect echoes from different beam lines.<br />
Unlike B-mode imaging that makes use of the backscattered echoes from tissues, color<br />
flow imaging relies on the blood scatterers inside the circulation as the primary signal targets.<br />
The strengths of backscattered blood echoes, however, are at least 10-20 dB lower than the<br />
backscattering strengths of tissues (Jensen 1996, Sec. 2.2). To compensate <strong>for</strong> this significant<br />
drop in the echo strength, it is not suitable to simply increase the intensity of the transmit pulses<br />
because of tissue heating concerns and safety regulations. Instead, the transmit pulses are made<br />
longer in duration than the ones used in B-mode scans so that the blood echo sensitivity and the<br />
blood-signal-to-noise ratio can be improved. Correspondingly, the transmit pulses <strong>for</strong> color flow<br />
imaging are often in the <strong>for</strong>m of multi-cycle sinusoidal bursts, and they are more narrowband in<br />
nature than the ones used <strong>for</strong> B-mode imaging. Note that the pulse duration T P can generally be<br />
approximated as follows:<br />
N<br />
cyc<br />
T<br />
P<br />
= , (1-4)<br />
f<br />
o<br />
† This statement is made in reference to the array probes used in current top-of-the-line scanners including Philips<br />
iU22, Siemens Sequoia 512, and GE LOGIQ 9.
Chapter 1. Introduction to <strong>Ultrasound</strong> <strong>Color</strong> <strong>Flow</strong> Imaging 12<br />
where N cyc denotes the number of sinusoidal cycles in the transmit pulse and f o is the pulse<br />
carrier frequency.<br />
Image Resolution and Penetration Depth<br />
The image resolution of color flow images, like B-mode images, is mainly defined by the<br />
type of transmit pulse used <strong>for</strong> the firings. In particular, the lateral resolution of the image<br />
depends on the lateral width of the beam profile generated from the transmit pulse, while the<br />
axial resolution depends on the duration of the transmit pulse. As described in ultrasound<br />
textbooks (e.g. see Cobbold 2006, Ch. 3), the lateral resolution l res corresponding to various<br />
focused beam patterns can be approximated as follows if the aperture width W is much greater<br />
than the pulse center wavelength λ o (i.e. W >> λ o ):<br />
F<br />
l =<br />
Fc<br />
o o<br />
res<br />
≈ λ , (1-5a)<br />
W Wf<br />
o<br />
where F is the beam focal length and it is well-known from wave physics that λ o equals to c o /f o .<br />
On the other hand, the axial resolution d res corresponding to a multi-cycle sinusoidal pulse can be<br />
approximated by:<br />
c T co<br />
N<br />
cyc<br />
≈ , (1-5b)<br />
2 2 f<br />
o P<br />
d<br />
res<br />
=<br />
where T P is as defined in (1-4) and the factor of one-half is included to account <strong>for</strong> the round-trip<br />
distance between the beam origin and a particular depth. As seen in (1-5a) and (1-5b), the lateral<br />
and axial resolutions both become finer at higher ultrasound frequencies. In general, the image<br />
resolution is between 0.1-1.0 mm <strong>for</strong> ultrasound frequencies in the operating range of array<br />
transducers (between 1-17 MHz). This resolution is sufficient to study the flow dynamics in<br />
large blood vessels and body organs, but a higher resolution is needed to study blood flow in the<br />
microcirculation where vessel diameters are typically less than 50 μm.<br />
Unlike the image resolution, the penetration depth of a pulse firing decreases at higher<br />
frequencies because the acoustic attenuation in tissues increases roughly as a linear function of<br />
the ultrasound frequency. In other words, any gain in the image resolution via an increase in the<br />
ultrasound frequency is accompanied by a concomitant decrease in the penetration depth. For a<br />
given attenuation factor α dB , the penetration depth d pen can be approximated as follows provided<br />
that the acoustic intensity decays exponentially with depth (Cobbold 2006, Sec. 1.8):<br />
o
Chapter 1. Introduction to <strong>Ultrasound</strong> <strong>Color</strong> <strong>Flow</strong> Imaging 13<br />
d<br />
pen<br />
I<br />
pen(dB)<br />
≈ , (1-6)<br />
α f<br />
where I pen(dB) is the intensity attenuation (in dB) at the penetration depth. In soft tissues and<br />
blood, the attenuation coefficient usually ranges between 0.2-0.7 dB/(cm⋅MHz) (Evans and<br />
McDicken 2000, pp. 38), and thus the 20-dB penetration depth is between 1-20 cm <strong>for</strong><br />
ultrasound frequencies less than 20 MHz.<br />
Velocity Resolution and Aliasing Limit<br />
The pulse-echo data acquisition process in color flow imaging can be considered as a<br />
sampling mechanism in which a single Doppler data sample is collected <strong>for</strong> each sample volume<br />
during each pulse firing. Owing to this sampling mechanism, the resulting flow estimates are<br />
inherently limited to a finite spectral resolution if they are derived from analyzing the inter-pulse<br />
phase changes. From sampling theory, it is well-known that the spectral resolution inversely<br />
depends on the ensemble period or, equivalently, the product of ensemble size and sampling<br />
period. Thus, the Doppler spectral resolution f D(res) can be approximated as follows:<br />
dB<br />
D<br />
o<br />
f = 1<br />
D(res)<br />
N T<br />
, (1-7a)<br />
where N D and T PRI are the Doppler ensemble size and the pulse repetition interval respectively<br />
(parameters to be described in the next subsection). Using the Doppler equation, the expression<br />
<strong>for</strong> Doppler resolution can be translated into the following expression <strong>for</strong> velocity resolution v res :<br />
PRI<br />
co<br />
v<br />
res<br />
= . (1-7b)<br />
f N T cosθ<br />
2<br />
o D PRI<br />
As seen in (1-7b), an increase in the ensemble period or ultrasound frequency can result in an<br />
improved velocity resolution, which is often important <strong>for</strong> studying slow flow dynamics.<br />
Amongst these factors, it is more suitable to increase the ultrasound frequency when imaging<br />
microcirculatory flow near the skin surface because a gain in both velocity resolution and image<br />
resolution can be achieved. On the other hand, it is necessary to use a longer ensemble period<br />
when imaging low-velocity flow at greater depths because high-frequency ultrasound cannot<br />
penetrate well into tissues.<br />
Another issue arising from the sampled nature of Doppler data is related to the maximum<br />
detectable frequency. From sampling theory, it is well-established that the maximum detectable
Chapter 1. Introduction to <strong>Ultrasound</strong> <strong>Color</strong> <strong>Flow</strong> Imaging 14<br />
frequency without aliasing is equal to half the sampling rate. Hence, the velocity aliasing limit<br />
v alias has the following <strong>for</strong>m (Cobbold 2006, Sec. 10.2):<br />
co<br />
v<br />
v<br />
alias<br />
=<br />
=<br />
4 f T cosθ<br />
o<br />
PRI<br />
res<br />
N<br />
2<br />
D<br />
, (1-8)<br />
where (1-7b) has been substituted into the above. As seen in (1-8), there is an inherent tradeoff<br />
in the maximum detectable velocity whenever a finer velocity resolution is achieved by<br />
increasing the ultrasound frequency or the pulse repetition interval. In order to detect flow<br />
beyond the velocity aliasing limit, it is necessary to derive the flow estimates from analyzing the<br />
inter-pulse time shifts instead of the inter-pulse phase changes. Nevertheless, the time-shiftbased<br />
estimation approach is often not necessary in color flow imaging because the presence of<br />
aliasing can actually enhance the visualization of blood flow dynamics (Tamura et al. 1991).<br />
Pulse Repetition Interval and Doppler Ensemble Size<br />
In color flow imaging, the range of pulse repetition intervals that can be used <strong>for</strong> data<br />
acquisition is bounded by two factors: real-time imaging requirements and imaging depth limits.<br />
In particular, the maximum pulse repetition interval mainly depends on real-time imaging<br />
constraints because all the pulse echoes from every beam line need to be acquired within a frame<br />
period that corresponds to a real-time frame rate (i.e. at least 5-10 Hz). On the contrary, the<br />
minimum pulse repetition interval depends primarily on the maximum imaging depth (which<br />
may not necessarily be the penetration depth) because sufficient time between firings is needed<br />
to collect pulse echoes from the field of view. <strong>Based</strong> on the two constraints, the pulse repetition<br />
interval T PRI must be defined within the following range <strong>for</strong> a given maximum depth d max ,<br />
acoustic speed c o , frame period T F , number of beam lines N L , and Doppler ensemble size N D :<br />
2d<br />
c<br />
T<br />
max<br />
F<br />
≤ TPRI<br />
≤ . (1-9)<br />
o<br />
N<br />
L<br />
N<br />
D<br />
In the above expression, the left side is essentially the round-trip propagation time with respect to<br />
the maximum imaging depth, and the right side is the maximum allowable time between each<br />
firing in order <strong>for</strong> all the pulse echoes to be acquired.<br />
Amongst the three dependent factors seen in the upper limit of (1-9), the Doppler<br />
ensemble size is the only factor that can be modified <strong>for</strong> a given imaging dimension and frame<br />
rate. Typically, <strong>for</strong> a field of view with 30-60 beam lines, the ensemble size available <strong>for</strong> each
Chapter 1. Introduction to <strong>Ultrasound</strong> <strong>Color</strong> <strong>Flow</strong> Imaging 15<br />
Doppler sample volume is limited to fewer than 20 samples so that the pulse repetition interval<br />
can be set to a value that satisfies both real-time imaging constraints and imaging depth limits.<br />
Note that, as mentioned earlier, the Doppler ensemble size also has an effect on the Doppler<br />
spectral resolution. Hence, one way to improve the Doppler spectral resolution is to use a larger<br />
Doppler ensemble size provided that the pulse repetition interval can still satisfy the upper limit<br />
of (1-9).<br />
Parameter Optimization<br />
Optimization of the system parameters <strong>for</strong> color flow imaging is generally a difficult task<br />
because of the various imaging considerations that need to be accounted <strong>for</strong>. Hence, solutions to<br />
this optimization problem are often non-trivial and involve many degrees of freedom. In order to<br />
simplify the design considerations, commercial systems often start with a few user-based<br />
acquisition parameters and then optimize the other ones internally. For instance, most systems<br />
allow the user to adjust parameters such as pulse repetition interval, ultrasound frequency, and<br />
field of view dimensions; from these parameters, the system then internally selects other<br />
parameters such as pulse duration, focal length, aperture width, frame rate, and Doppler<br />
ensemble size (i.e. the number of firings per beam line). The effects of different user-defined<br />
data acquisition parameters on the other imaging parameters are summarized in Table 1-1. In<br />
this thesis, we shall focus our ef<strong>for</strong>t on addressing the effects of Doppler ensemble size and pulse<br />
repetition interval on the per<strong>for</strong>mance of color flow signal processors. As will be noted in<br />
Section 1.5, these two parameters often have a significant impact on the efficacy of the Doppler<br />
signal processing strategies.<br />
1.4.3 Advanced <strong>Flow</strong> Imaging Approaches<br />
Three-Dimensional Mapping<br />
Over the past decade, various advanced data acquisition approaches have been proposed<br />
to expand upon the flow in<strong>for</strong>mation provided in color flow images. One representative example<br />
of these advanced approaches is three-dimensional (3D) color flow imaging (Fenster et al. 2001).<br />
The main motivation <strong>for</strong> the development of 3D flow imaging is that conventional color flow<br />
images can only provide two-dimensional (2D) mono-planar flow maps along the transducer’s<br />
field of view. Such limitation may mask out important flow in<strong>for</strong>mation under disturbed flow
Chapter 1. Introduction to <strong>Ultrasound</strong> <strong>Color</strong> <strong>Flow</strong> Imaging 16<br />
Table 1-1. Effects of User-Defined Data Acquisition Parameters<br />
Parameter<br />
Increased ultrasound frequency<br />
(↑ f o )<br />
Increased pulse repetition interval<br />
(↑ T PRI )<br />
Increased lateral field of view<br />
Increased axial field of view<br />
Effects<br />
• Finer image resolution (↓ l res , d res )<br />
• Lower penetration depth (↓ d pen )<br />
• Finer velocity resolution (↓ v res )<br />
• Lower velocity aliasing limit (↓ v alias )<br />
• Finer velocity resolution (↓ v res )<br />
• Lower velocity aliasing limit (↓ v alias )<br />
• More beam lines (↑ N L )<br />
• Longer frame period and/or smaller Doppler<br />
ensemble size (↑ T F , ↓ N D )<br />
• Coarser velocity resolution (if Doppler<br />
ensemble size is decreased) (↑ v res )<br />
• Increased maximum imaging depth (↑ d max )<br />
• Longer pulse duration (to improve blood<br />
echo sensitvity) (↑ T P )<br />
• Coarser axial resolution (↑ d res )<br />
scenarios, as a mono-planar flow map inherently lacks the ability to reveal the 3D nature of<br />
turbulent flow dynamics. To overcome this difficulty, some research studies have attempted to<br />
use special transducer setups to acquire multiple slices of mono-planar color flow images and<br />
stack them together as a 3D volumetric display. For instance, Picot et al. (1993) have produced<br />
3D flow maps by physically moving the transducer to acquire 2D flow images along different<br />
scan planes. Similar flow images can also be generated through electronic scan plane steering<br />
with 2D array transducers (von Ramm et al. 1991). This latter approach appears to be more<br />
robust because it is more capable of acquiring 3D flow maps in real time.<br />
Vector <strong>Flow</strong> Imaging<br />
Vector-based flow imaging is another advanced scheme that is useful in vascular flow<br />
studies. As reviewed by Dunmire et al. (2000), the objective of vector flow imaging is to obtain<br />
consistent color flow estimates without prior knowledge of the beam-flow angle. This approach<br />
is beneficial to the study of complex flow patterns because the beam-flow angle may vary over<br />
the imaged vessel, and thus the velocity estimates may be less accurate if a fixed beam-flow<br />
angle is assumed. In terms of its principles, most vector flow imaging methods work by
Chapter 1. Introduction to <strong>Ultrasound</strong> <strong>Color</strong> <strong>Flow</strong> Imaging 17<br />
combining the Doppler frequency shifts estimated from different beam directions based on<br />
geometric reconstruction techniques. This multi-beam configuration can be implemented<br />
physically using two different approaches. The first is to use multiple transducers <strong>for</strong> excitation<br />
and reception, such as that used in the study by Overbeck et al. (1992). The second is to use a<br />
single array transducer that has beam steering capabilities, such as the one used by Maniatis et al.<br />
(1994). <strong>Based</strong> on the second approach, Capineri et al. (2002) have developed an experimental<br />
vector flow imaging system that can provide angle-independent flow maps in real time. Aside<br />
from the use of multiple-beam setups, Jensen and Munk (1998) have proposed to obtain vector<br />
flow estimates using a single ultrasound beam that has a modified beam<strong>for</strong>ming pattern. This<br />
single-beam vector flow imaging approach has recently been implemented on an experimental<br />
scanner (Udesen and Jensen 2006).<br />
High-Frequency <strong>Flow</strong> Imaging<br />
A third advanced imaging scheme of clinical interest is high-frequency color flow<br />
imaging that involves the use of ultrasound waves in the 30-100 MHz range (Foster et al. 2000).<br />
The primary advantage of using high-frequency ultrasound <strong>for</strong> color flow imaging is its superior<br />
spatial resolution as compared to conventional imaging that is typically done at frequencies<br />
below 15 MHz. With this improved spatial resolution, the resulting color flow images can<br />
provide visualization of the flow dynamics inside the microvasculature and hence obtain new<br />
insights on the physiology of microcirculation. In terms of its development, Ferrara et al. (1996)<br />
have developed a 38 MHz experimental flow estimation system, while Christopher et al. (1997)<br />
have reported on a similar apparatus that operates at 50 MHz. More recently, Goertz et al.<br />
(2000) have expanded upon the previous designs by building a 20-100 MHz color flow imaging<br />
system prototype. Kruse and Ferrara (2002) have also reported the development of a 25-38 MHz<br />
color flow imaging system that is based on sweep-scanning principles.<br />
Microbubble Contrast Agents<br />
The use of microbubble contrast agents <strong>for</strong> flow imaging is another advanced technique<br />
that has many potential clinical applications (Frinking et al. 2000). The aim of this technique is<br />
to facilitate better flow detection through the nonlinear scattering properties of microbubbles.<br />
Specifically, flow detection can be improved by injecting microbubbles into the vasculature and<br />
subsequently processing the fundamental and higher-harmonic echoes returned from these
Chapter 1. Introduction to <strong>Ultrasound</strong> <strong>Color</strong> <strong>Flow</strong> Imaging 18<br />
contrast agents during their passage through the imaged vessel. Microbubbles and harmonic<br />
flow imaging are particularly useful in the visualization of flow dynamics inside small vessels<br />
where echoes from surrounding tissues tend to mask out the ones from red blood cell scatterers.<br />
Regarding the efficacy of microbubbles, several flow estimation studies (e.g. Schrope and<br />
Newhouse 1993, Chang et al. 1996) have experimentally evaluated the strength of harmonic<br />
echoes at different contrast agent concentrations. From these experimental findings, Hope<br />
Simpson et al. (1999) have subsequently proposed the use of a pulse inversion Doppler imaging<br />
scheme to further amplify the second harmonic echoes returned from microbubbles. This<br />
scheme works by transmitting a pulse as well as its negated <strong>for</strong>m in sequence and summing the<br />
two complementary pulse echoes to cancel out the odd-harmonic signals while retaining the<br />
even-harmonic ones. With the pulse inversion method, Bruce et al. (2004) have reported<br />
significant improvements in the Doppler signal quality when microbubbles are used during data<br />
acquisition.<br />
Additional Remarks<br />
Even though their data acquisition principles are more advanced, the new flow imaging<br />
approaches described above (perhaps with the exception of pulse inversion Doppler) still make<br />
use of the same signal processing strategies as used in conventional color flow imaging. As<br />
such, their usefulness to vascular flow studies inherently depends on whether the clutter filters<br />
and flow estimators can per<strong>for</strong>m well under various imaging conditions. Because of this<br />
dependence, it is important <strong>for</strong> the clutter filters to be able to adequately suppress low-frequency<br />
clutter and <strong>for</strong> the flow estimators to be able to obtain consistent velocity and power estimates.<br />
Otherwise, there may be noticeable errors in the additional flow in<strong>for</strong>mation provided by the<br />
advanced flow imaging approaches.<br />
1.5 <strong>Signal</strong> <strong>Processing</strong> Challenges in <strong>Color</strong> <strong>Flow</strong> Imaging<br />
1.5.1 Background Considerations<br />
As pointed out earlier, the primary signal targets <strong>for</strong> ultrasound-based flow estimation<br />
schemes are the red blood cells moving inside the vessel. As such, in the ideal case, the acquired<br />
Doppler signal <strong>for</strong> a sample volume (or map pixel location) encompassing blood flow should<br />
only comprise of non-zero-frequency components that arise from the inter-pulse phase changes
Chapter 1. Introduction to <strong>Ultrasound</strong> <strong>Color</strong> <strong>Flow</strong> Imaging 19<br />
Clutter<br />
Spectrum<br />
Filter Response<br />
Noise Floor<br />
0<br />
Blood<br />
Spectrum<br />
f D<br />
Fig. 1-4. Illustration of the typical spectral contents in a Doppler signal be<strong>for</strong>e digital<br />
sampling. Prior to flow estimation, it is often necessary to use a highpass filter to suppress<br />
the high-energy and low-frequency clutter in the Doppler signal.<br />
induced by moving blood scatterers. However, since the vessel is generally surrounded by other<br />
scattering sources like tissues and vessel walls, the Doppler signal may also contain undesired<br />
components that are due to reverberations from non-blood scatterers. In this thesis, these<br />
undesired Doppler signal components will be referred to as clutter, and they should be<br />
distinguished from the background white noise that naturally arises from random fluctuations in<br />
the data acquisition electronics. It is worth pointing out that, even though a sample volume may<br />
be nominally positioned inside the blood vessel, clutter may still be seen in the resulting Doppler<br />
signal because the diffractive nature of ultrasound beams inherently gives rise to beam sidelobes<br />
(sometimes referred to as leakage lobes) that encompass non-vessel regions outside the nominal<br />
sample volume range.<br />
In contrast to the moving blood scatterers, tissues and vessel walls tend to exhibit a<br />
significantly smaller degree of movement, and thus they often induce smaller degrees of interpulse<br />
phase changes. <strong>Based</strong> on this premise, the corresponding Doppler clutter is usually lowfrequency<br />
in nature and does not overlap with the blood spectral components on the Doppler<br />
frequency axis (as illustrated in Fig. 1-4). Nevertheless, since tissues have a stronger scattering<br />
strength than blood, the magnitude of Doppler clutter is usually much higher than that of the<br />
desired blood Doppler echoes. In fact, as suggested in various studies (Kadi and Loupas 1995,<br />
Torp 1997, Bjaerum et al. 2002a), the clutter strength can be as much as 60-80 dB greater in<br />
magnitude depending on the dynamic data range supported by the scanner. Consequently, their<br />
presence in the Doppler signal makes it more challenging to extract accurate blood flow<br />
in<strong>for</strong>mation from the acquired data.
Chapter 1. Introduction to <strong>Ultrasound</strong> <strong>Color</strong> <strong>Flow</strong> Imaging 20<br />
1.5.2 Challenges in Clutter Suppression<br />
Doppler Ensemble Size Limitations<br />
As already shown in Fig. 1-3, the first step in color flow signal processing is to apply a<br />
highpass filter to each Doppler ensemble so that any low-frequency clutter can be adequately<br />
suppressed prior to the computation of blood flow in<strong>for</strong>mation. However, this clutter filtering<br />
operation is a rather challenging task in color flow imaging because less than 20 samples of<br />
Doppler data are usually available <strong>for</strong> processing owing to real-time imaging constraints (see<br />
Section 1.4.2). With such a small data size, the clutter filter’s transient characteristics tend to<br />
dominate the filter output and in turn reduce the efficacy of clutter suppression.<br />
To effectively remove clutter from Doppler signals that have small ensemble sizes,<br />
studies have considered various ways of minimizing the transient response of a clutter filter. For<br />
instance, Bjaerum et al. (2002a) have reported that the transient effects of finite impulse response<br />
(FIR) filters can be mitigated by implementing the minimum-phase equivalent of this filter. In<br />
contrast, <strong>for</strong> infinite impulse response (IIR) filters, studies have shown that filter initialization<br />
schemes like step initialization, projection initialization, and exponential initialization can be<br />
used to reduce the filter’s transient characteristics (Peterson et al. 1994, Kadi and Loupas 1995).<br />
Note that, as another way to circumvent filter transient problems, it is also possible to use<br />
polynomial regression filters to suppress clutter in Doppler ensembles (Hoeks et al. 1991, Kadi<br />
and Loupas 1995, Torp 1997). This type of filter, whose response is time-variant in nature,<br />
works by finding the least-squares fit of the Doppler signal onto a polynomial subspace and<br />
subtracting this fit from the Doppler signal. In examining the efficacy of these filtering<br />
approaches, Bjaerum et al. (2002a) have concluded that the clutter suppression per<strong>for</strong>mances of<br />
the projection-initialized IIR filter and the polynomial regression filter are generally better than<br />
the other ones because of their time-variant filter characteristics. Nevertheless, it should be<br />
noted that these filters all require an appropriate selection of the filter stopband in order to<br />
achieve adequate clutter suppression. To address this issue, Yoo et al. (2003) have described an<br />
adaptive approach to specify the filter stopband based on the mean clutter frequency.<br />
Wideband Clutter Problems<br />
Another challenge that affects the efficacy of clutter suppression is related to the use of a<br />
long ensemble period <strong>for</strong> pulse echo data acquisition. As discussed in Section 1.4.2, a longer
Chapter 1. Introduction to <strong>Ultrasound</strong> <strong>Color</strong> <strong>Flow</strong> Imaging 21<br />
ensemble period sometimes needs to be used to detect low flow velocities in the deeper sample<br />
volumes. However, one major problem of lengthening the ensemble period <strong>for</strong> data acquisition<br />
is that tissues may undergo substantial motion over this extended period, and consequently the<br />
Doppler signal may contain clutter echoes that are more wideband in nature and that are not<br />
centered around zero frequency (Heimdal and Torp 1997). To suppress clutter in Doppler<br />
signals acquired over a lengthened period, it is necessary to widen the stopband of the clutter<br />
filter so that the possible wideband nature of clutter can be properly accounted <strong>for</strong>. Nevertheless,<br />
a widened filter stopband may concomitantly suppress a substantial portion of the blood echoes.<br />
To improve the suppression of wideband clutter, some studies have considered the use of<br />
more advanced filtering approaches. For instance, Thomas and Hall (1994) as well as Brands et<br />
al. (1995) have proposed an approach that first involves the downmixing of the Doppler signal<br />
with the mean clutter frequency estimate prior to the clutter filtering operation. As a result of<br />
this downmixing step, the clutter spectral components are shifted toward zero Doppler<br />
frequency, and hence it is more likely that the clutter echoes can be removed by a conventional<br />
clutter filter that has a narrower stopband. Another way of suppressing wideband clutter, as<br />
described by Ledoux et al. (1997), is to make use of a principal component analysis framework<br />
whereby the clutter echoes are adaptively filtered out by removing the most dominant<br />
components in the singular value decomposition of the Doppler data at various depths. Using the<br />
same principal subspace concept, Bjaerum et al. (2002b) have reported an eigen-based regression<br />
filter that is based on an eigendecomposition of the Doppler correlation statistics. Similarly,<br />
Kruse and Ferrara (2002) have developed an eigen-filtering strategy <strong>for</strong> use in swept-scan flow<br />
imaging applications, while Kargel et al. (2003) have applied the same approach to strain-flow<br />
imaging studies. An extension to the eigen-regression filtering approach has recently been<br />
considered by Gallippi et al. (2003), who made use of an independent component analysis<br />
framework to suppress Doppler clutter in acoustic radiation <strong>for</strong>ce imaging studies. As compared<br />
to the clutter-downmixing filter, the eigen-regression filter seems to be more capable of<br />
suppressing Doppler clutter while preserving blood flow echoes. However, the primary<br />
challenge of using this type of filter is that the Doppler correlation matrix often needs to be<br />
estimated from multiple Doppler signal snapshots that are statistically stationary. As such, this<br />
filter may be challenging to implement in some vascular imaging studies where it is difficult to<br />
segment out regions with similar Doppler data statistics. Another limitation of the eigen-
Chapter 1. Introduction to <strong>Ultrasound</strong> <strong>Color</strong> <strong>Flow</strong> Imaging 22<br />
regression filter is that its efficacy is rather sensitive to the choice of the clutter eigen-space<br />
dimension (as analogous to the filter stopband of FIR/IIR filters) owing to the filter’s adaptive<br />
nature. Hence, it is necessary to develop an effective algorithm <strong>for</strong> this filter to select the clutter<br />
eigen-space dimension.<br />
1.5.3 Challenges in Velocity Estimation<br />
Carrier Frequency Variations<br />
Given the filtered Doppler signal, one way of estimating the mean flow velocity is to<br />
compute the mean Doppler frequency from the Doppler spectral moments and apply the Doppler<br />
equation to convert the frequency estimate into a velocity value. <strong>Based</strong> on this notion, Kasai et<br />
al. (1985) have reported a closed-<strong>for</strong>m mean velocity estimator that works by analyzing the lagone<br />
autocorrelation of the filtered Doppler signal. This estimator has low computational<br />
demands, and it is generally unbiased in the presence of white noise. However, its mean Doppler<br />
frequency estimates may be biased when there are variations in the carrier frequency of the<br />
returned echoes simply because the pulse carrier frequency is one of the governing factors in the<br />
Doppler equation (Ferrara et al. 1992). As pointed out in a few studies (Round and Bates 1987,<br />
Embree and O’Brien 1990), attenuation and scattering are the two primary distortion<br />
mechanisms of the pulse carrier frequency. They tend to bring about a gradual shift in the pulse<br />
carrier frequency as depth increases, and the shift magnitude per unit depth is mainly dependent<br />
on the bandwidth of the transmit pulse. Fortunately, this problem is usually of less concern in<br />
color flow imaging because narrowband transmit pulses typically need to be used in order to<br />
improve the blood echo sensitivity and maintain adequate blood-signal-to-noise ratios (as noted<br />
in Section 1.4.2). Nonetheless, variations in the carrier frequency can become a significant<br />
velocity biasing source when wideband transmit pulses are used <strong>for</strong> data acquisition.<br />
To account <strong>for</strong> biases due to carrier frequency variations, one straight<strong>for</strong>ward solution is<br />
to estimate the carrier frequency shifting rate as a function of depth and apply a correction factor<br />
to the mean Doppler frequency estimates based on the shifting rate. Alternatively, a number of<br />
some studies have considered the development of more advanced flow estimation strategies to<br />
account <strong>for</strong> carrier frequency variations. For instance, Loupas et al. (1995a) have reported a 2D<br />
autocorrelation estimator that can compute both the mean Doppler frequency and the mean echo<br />
carrier frequency based on the Doppler signals of adjacent sample volumes along the same beam
Chapter 1. Introduction to <strong>Ultrasound</strong> <strong>Color</strong> <strong>Flow</strong> Imaging 23<br />
line. On the other hand, some studies have proposed the use of target tracking methods such as<br />
cross-correlation (Bonnefous and Pesque 1986) and maximum likelihood estimation (Ferrara and<br />
Algazi 1991a, Alam and Parker 1995). These approaches work by analyzing the inter-pulse<br />
displacement of the complex echo envelope, and thus they are inherently unaffected by<br />
variations in the carrier frequency. However, these correlation-based estimation approaches<br />
remain susceptible to biases originating from clutter filter distortions.<br />
Clutter Filter Distortions<br />
The primary side effect of using a filter to suppress low-frequency clutter is that the<br />
blood echoes may be concomitantly distorted due to non-idealities in the filter’s frequency<br />
response (Willemetz et al. 1989, Nowicki et al 1990, Tysoe and Evans 1995). The distortions<br />
can be particularly severe if the blood echoes’ Doppler spectral components are located near the<br />
filter’s cutoff frequency and its transition region. As a result of these potential distortions, there<br />
may be a bias in the mean frequency estimates computed from the filtered Doppler signal. Note<br />
that, at low blood-signal-to-noise ratios, the biasing problem becomes worse because clutter filter<br />
also distorts the background white noise and thereby adds further bias to the flow estimates. As<br />
pointed out by Rajaonah et al. (1994), post facto bias correction may be applied to the mean<br />
frequency estimates if a time-invariant filter (e.g. an FIR filter or non-initialized IIR filter) is<br />
used to suppress clutter. Nevertheless, <strong>for</strong> time-variant filters like the ones discussed earlier, this<br />
method itself may not be adequate to account <strong>for</strong> the filter biases.<br />
To avoid biases due to clutter filtering, some studies have considered the development of<br />
parametric spectral estimators that can compute the principal Doppler frequency of blood echoes<br />
in the presence of clutter. For instance, Ahn and Park (1991) have applied the autoregressive<br />
(AR) modeling method originally reviewed by Vaitkus and Cobbold (1988) to simultaneously<br />
extract both the principal clutter and blood Doppler frequencies from the raw (i.e. unfiltered)<br />
Doppler signal. In their approach, the raw Doppler signal is modeled as a two-pole signal where<br />
one of the signal poles corresponds to clutter and the other corresponds to blood signals. The<br />
signal poles are then solved to obtain the principal frequency estimate of clutter and blood<br />
signals. In contrast, Vaitkus and Cobbold (1998) have proposed the use of an eigen-analysis<br />
framework called “multiple signal classification” (MUSIC) <strong>for</strong> flow estimation without prior<br />
clutter filtering. This approach, which was originally examined by Allam and Greenleaf (1996),
Chapter 1. Introduction to <strong>Ultrasound</strong> <strong>Color</strong> <strong>Flow</strong> Imaging 24<br />
models the Doppler signal as a summation of two principal harmonics, and it solves <strong>for</strong> the<br />
harmonics from the eigen-decomposition of the Doppler signal’s correlation matrix statistics.<br />
The primary limitation of the MUSIC approach and the AR method, however, is that the Doppler<br />
signal cannot always be modeled as a two-pole or two-harmonic signal (especially when the<br />
clutter is wideband in nature). As such, an adaptive model order selection algorithm is needed in<br />
order <strong>for</strong> these two estimators to be effective in general.<br />
1.6 Concluding Remarks<br />
In clinical diagnostic procedures, ultrasound color flow imaging is primarily used as a<br />
qualitative way of visualizing blood flow dynamics. Nevertheless, it is still important to provide<br />
accurate flow in<strong>for</strong>mation in color flow images so that the risk of misdiagnosis can be reduced.<br />
In order to maintain the reliability of in<strong>for</strong>mation provided in color flow images, however, there<br />
are a few signal processing challenges that first need to be overcome during clutter suppression<br />
and flow estimation. As discussed in the previous section, the major challenges encountered<br />
during clutter suppression are mainly associated with the lack of abundant Doppler samples<br />
available <strong>for</strong> data processing as well as the possible presence of wideband clutter when using<br />
longer data acquisition periods. On the other hand, the primary challenge <strong>for</strong> flow estimation is<br />
related to the flow signal distortions originating from clutter suppression. To address these<br />
challenges, the subsequent chapters of this thesis will present new eigen-based color flow signal<br />
processing methods that are based on the eigen-space properties of two matrix <strong>for</strong>ms: the Hankel<br />
matrix and the matrix pencil. In particular, the remaining chapters have been organized as<br />
follows:<br />
• Chapter 2 provides a theoretical treatment of the principles behind the Hankel-SVD filter<br />
and describes this filter’s ability to suppress clutter in the Doppler signal;<br />
• Chapter 3 presents the theory and principles of the Matrix Pencil estimation framework and<br />
discusses this approach’s application in color flow data processing;<br />
• Chapter 4 describes how the Hankel-SVD filter and the Matrix Pencil estimator can be<br />
combined into a single-module color flow signal processor;<br />
• Chapter 5 summarizes the contributions of this thesis study and discusses some possible<br />
future directions in this research area.
CHAPTER 2<br />
Clutter Suppression in <strong>Color</strong> <strong>Flow</strong> Imaging:<br />
A Hankel-SVD Approach<br />
2.1 Chapter Overview<br />
As described in the introductory chapter, there are two problems that significantly<br />
increase the difficulty of clutter suppression in color flow imaging. The first is related to the<br />
small Doppler ensemble size (usually less than 20 samples) that is available <strong>for</strong> flow estimation<br />
because of the need to satisfy real-time imaging constraints. Due to such a small sample size, the<br />
filtered Doppler signal can be significantly affected by the transient characteristics of the clutter<br />
filter. The second problem is concerned with the longer data ensemble period (on the order of<br />
100 ms) that is sometimes required to detect slow flow dynamics at greater depths. Since tissues<br />
may undergo substantial motion over this extended period, the clutter component of the Doppler<br />
signal may be more wideband in nature and may not be centered around zero frequency.<br />
In attempt to address the problems in Doppler clutter suppression, this chapter presents<br />
the development of a new eigen-based, adaptive clutter filter design <strong>for</strong> color flow imaging<br />
applications. This new filter is based on a singular value decomposition (SVD) analysis of the<br />
Hankel matrix constructed from the Doppler signal of each sample volume, and hence it will be<br />
referred to as the Hankel-SVD filter from hereon. To <strong>for</strong>mulate discussion on the Hankel-SVD<br />
filter, the rest of this chapter has been organized as follows:<br />
• Section 2.2 surveys the highlights of existing clutter filter designs and discusses their<br />
principles used to suppress Doppler clutter;<br />
• Section 2.3 presents the theoretical principles of the Hankel-SVD filter, describes its<br />
adaptability in suppressing wideband clutter, and discusses the computational considerations<br />
involved with the implementation of this filter;<br />
• Section 2.4 describes the details of Doppler signal synthesis method used to study the new<br />
filter’s flow detection per<strong>for</strong>mance;<br />
- 25 -
Chapter 2. Clutter Suppression in <strong>Color</strong> <strong>Flow</strong> Imaging: A Hankel-SVD Approach 26<br />
• Section 2.5 presents the analysis results <strong>for</strong> the Hankel-SVD filter’s per<strong>for</strong>mance and<br />
compares them with the ones obtained from other types of clutter filters;<br />
• Section 2.6 summarizes the highlights of the new filter and further discusses some practical<br />
considerations on the new filter design.<br />
2.2 Existing Clutter Filter Designs<br />
2.2.1 Non-Adaptive Filters<br />
Background Considerations<br />
To facilitate description of clutter filter designs, we first consider the Doppler signal<br />
model <strong>for</strong> a sample volume (or map pixel location) that encompasses blood flow. In this case,<br />
the Doppler signal generally comprises low-frequency tissue clutter, high-frequency blood<br />
echoes, and random white noise. For a given Doppler ensemble size N D , these signal<br />
components can be expressed in the following vector <strong>for</strong>m:<br />
T<br />
[ ( 0), x(1),<br />
, x(<br />
N −1<br />
] = b + c w<br />
x = x K + (in vasculature), (2-1a)<br />
D<br />
)<br />
where x(n) is the n th Doppler data sample, while x, b, c, and w are vectors of length N D <strong>for</strong> the<br />
Doppler signal, blood echoes, clutter, and white noise respectively. On the other hand, <strong>for</strong> a<br />
sample volume located in the tissue region, its Doppler signal should only comprise of tissue<br />
clutter and random white noise (since blood flow is not present). Correspondingly, the signal<br />
equation <strong>for</strong> such a sample volume is given by:<br />
T<br />
[ ( 0), x(1),<br />
, x(<br />
N −1<br />
] = c w<br />
x = x K + (in tissue). (2-1b)<br />
D<br />
)<br />
To study whether blood signal is present in a Doppler ensemble, it is necessary to first<br />
use a highpass filter to suppress the low-frequency clutter whose strength is usually much higher<br />
(greater than 20 dB) than the other Doppler signal components. From a vector space perspective,<br />
this filtering task can be considered as the application of a linear matrix operator on the Doppler<br />
signal vector. Thus, the filtered Doppler signal vector y can be written as follows:<br />
T<br />
[ ( 0), y(1),<br />
, y(<br />
N −1<br />
] Fx<br />
y = y K = , (2-2)<br />
D<br />
)<br />
where y(n) is the n th filtered sample and F is an N D ×N D filter matrix. Note that, depending on<br />
how the filter matrix is <strong>for</strong>med, the clutter filter may have a time-variant frequency response.<br />
Thus, as pointed out by Torp (1997), the filter’s frequency response may not simply correspond
Chapter 2. Clutter Suppression in <strong>Color</strong> <strong>Flow</strong> Imaging: A Hankel-SVD Approach 27<br />
to the Fourier trans<strong>for</strong>m of the filter’s impulse response. Nevertheless, it is still possible to<br />
numerically compute the frequency response of these time-variant clutter filters by finding the<br />
filter’s output power <strong>for</strong> different complex sinusoid inputs. <strong>Based</strong> on this notion, the filter<br />
response H(f D ) at a particular Doppler frequency can be expressed as:<br />
2 H H<br />
( f<br />
D<br />
) = Fv<br />
CS<br />
= v<br />
CSF<br />
Fv<br />
CS<br />
H , (2-3a)<br />
where v CS is the following complex sinusoid vector of length N D :<br />
CS<br />
j 2πf<br />
2 ( 2) 2 ( 1)<br />
[ ] T<br />
D j πf<br />
D ND<br />
−<br />
, j πf<br />
D ND<br />
−<br />
1, e , K,<br />
e e<br />
v =<br />
. (2-3b)<br />
FIR and IIR Filters<br />
It is well-known that the finite impulse response (FIR) filter and the infinite impulse<br />
response (IIR) filter are the two common types of linear filters used in signal processing. As<br />
described in various signal processing textbooks (e.g. Oppenheim et al. 1999, Ch. 6 and 7), the<br />
FIR filter is a non-recursive filter whose output samples only depend on the input data samples,<br />
and hence its impulse response has a finite duration. On the other hand, the IIR filter is a<br />
recursive filter whose output samples depend on both input data samples and past output, and<br />
correspondingly its impulse response carries on <strong>for</strong> an infinite time. For both types of filters, the<br />
filter order is essentially equivalent to the number of input samples that each filter output<br />
depends on. A higher filter order can yield a narrower transition band and a more uni<strong>for</strong>m<br />
passband in the filter’s steady-state frequency response, but it also brings about a lengthening of<br />
the filter’s transient response. Note that, <strong>for</strong> a K th -order IIR filter, its n th output sample can be<br />
expressed as follows:<br />
K<br />
∑<br />
k = 1<br />
K<br />
∑<br />
y( n)<br />
= − c y(<br />
n − k)<br />
+ b x(<br />
n − k)<br />
, (2-4a)<br />
k<br />
where c k and b k are the respective weights of past output samples and input samples used in the<br />
filter, and they can be found using approaches like Butterworth and Chebyshev design methods.<br />
The output <strong>for</strong> FIR filters essentially has the same <strong>for</strong>m as shown in (2-4a) except that all the<br />
weights of past output samples are set to zero (i.e. c k = 0). The input weights <strong>for</strong> FIR filters can<br />
be found using approaches such as the windowing method and the Parks-McClellan algorithm.<br />
From a state-space perspective, the output of both FIR and IIR filters can be considered<br />
as a joint contribution from the filter’s initial state and the input samples. In particular, as<br />
k = 0<br />
k
Chapter 2. Clutter Suppression in <strong>Color</strong> <strong>Flow</strong> Imaging: A Hankel-SVD Approach 28<br />
described by Bjaerum et al. (2002a), the filter output can be expressed in the following vector<br />
<strong>for</strong>m:<br />
y = Tv<br />
init<br />
+ Sx , (2-4b)<br />
where v init is the initial filter state vector of length K (i.e. the filter order), while T and S are<br />
respectively the transient filter matrix (of size N D ×K) and the steady-state filter matrix (of size<br />
N D ×N D ) whose entries depend on the filter weights c k and b k shown in (2-4a). For FIR filters, the<br />
initial filter state vector only affects the filter output <strong>for</strong> a finite duration; however, <strong>for</strong> IIR filters,<br />
this vector has an effect on the filter output all the time since the filter’s impulse response never<br />
vanishes. As such, the initial state vector should be defined in a way so that the filter’s transient<br />
effects can be mitigated. <strong>Based</strong> on this notion, a few studies have considered the use of timevariant<br />
initialization approaches such as step initialization and projection initialization (Peterson<br />
et al. 1994, Kadi and Loupas 1995). Recently, Bjaerum et al. (2002a) have shown that projection<br />
initialization is more effective because it gives a stopband frequency response that looks the most<br />
similar to the steady-state response. This initialization approach, as originally proposed by<br />
Chornoboy (1992), works by setting the initial state vector as the complement of the leastsquares<br />
fitting coefficients between the transient filter matrix and the steady-state filter output.<br />
In particular, the initial state vector <strong>for</strong> projection initialization can be expressed as follows:<br />
v<br />
H −1<br />
H<br />
+<br />
= −( T T T Sx = −T<br />
Sx , (2-5a)<br />
init<br />
)<br />
where the ‘+’ superscript denotes a pseudo-inverse operation (i.e the singular matrix equivalent<br />
of a square matrix inverse). By combining (2-4b) and (2-5a), the overall filter matrix with<br />
projection initialization can then be written in the following <strong>for</strong>m:<br />
+<br />
F = [ I − TT ] S , (2-5b)<br />
where I is an N D ×N D identity matrix. Fig. 2-1 shows an example of the frequency response <strong>for</strong><br />
projection-initialized IIR filters. This figure was computed by substituting (2-5b) into (2-3a),<br />
and it shows that the actual response of the projection-initialized IIR filter is suboptimal (but still<br />
close) to the steady-state filter response. Note that, besides the use of filter initialization<br />
schemes, minimum-phase filters can also be used to reduce transient filtering effects. Bjaerum et<br />
al. (2002a) have studied the use of minimum-phase FIR filters and have shown that they can<br />
improve the filtering per<strong>for</strong>mance.
Chapter 2. Clutter Suppression in <strong>Color</strong> <strong>Flow</strong> Imaging: A Hankel-SVD Approach 29<br />
(a) 5 th -order Filter<br />
(b) 8 th -order Filter<br />
Magnitude [dB]<br />
0<br />
-10<br />
-20<br />
-30<br />
-40<br />
-50<br />
-60<br />
Steady-State<br />
Actual<br />
0 50 100 150 200 250<br />
Doppler Frequency (f D<br />
) [Hz]<br />
0<br />
-10<br />
-20<br />
-30<br />
-40<br />
-50<br />
Steady-State<br />
-60<br />
Actual<br />
0 50 100 150 200 250<br />
Doppler Frequency (f D<br />
) [Hz]<br />
Fig. 2-1. Actual and steady-state filter responses of the projection-initialized IIR filter. The<br />
responses are shown <strong>for</strong>: (a) 5 th -order filter and N D =10; (b) 8 th -order filter and N D =20. The<br />
Doppler sampling frequency (i.e. 1/T PRI ) used in both plots was 500 Hz, and the IIR filters<br />
were designed using the Chebychev method with a nominal cutoff frequency of 50 Hz.<br />
Magnitude [dB]<br />
Regression Filters<br />
As another way of suppressing Doppler clutter, a regression filter may be used in place of<br />
FIR and IIR filters. This type of filter works by first modeling Doppler clutter as a series of<br />
curve shapes and then computing the least-squares fitting residual between the Doppler signal<br />
and the given clutter model. From a subspace perspective, such regressive fitting is equivalent to<br />
a least-squares projection of the Doppler signal vector onto the orthogonal complement of the<br />
clutter model subspace. As described in various algebra textbooks (e.g. Moon and Stirling 2000,<br />
Sec. 3.4.1), the least-squares projection of a vector x onto the orthogonal complement of a<br />
subspace matrix C is given by:<br />
H −1<br />
H<br />
+<br />
y = [ I − C(<br />
C C)<br />
C ] x = [ I − CC ] x . (2-6a)<br />
Consequently, the filter matrix <strong>for</strong> a regression filter can be expressed as follows:<br />
+<br />
F = I − CC . (2-6b)<br />
In terms of the curve shapes used <strong>for</strong> the clutter model, low-order polynomials are often suitable<br />
because clutter echoes are low frequency components in nature. Thus, polynomial basis vectors<br />
are typically used <strong>for</strong> the columns of the clutter subspace matrix. For a K th -order polynomial<br />
clutter model ĉ, the subspace matrix C is an N D ×(K+1) matrix of the following <strong>for</strong>m:<br />
⎡ | | | ⎤<br />
K<br />
cˆ = ⇔ =<br />
⎢<br />
⎥<br />
∑ χ<br />
kp<br />
k<br />
C<br />
⎢<br />
p<br />
0<br />
p1<br />
L p<br />
K<br />
⎥<br />
, (2-7)<br />
k = 0<br />
⎢⎣<br />
| | | ⎥⎦
Chapter 2. Clutter Suppression in <strong>Color</strong> <strong>Flow</strong> Imaging: A Hankel-SVD Approach 30<br />
where χ k and p k are the respective least-squares fitting coefficient and the length-N D basis vector<br />
<strong>for</strong> the k th -order polynomial (usually ranged between –1 and +1). <strong>Based</strong> on this notion, Hoeks et<br />
al. (1991) have considered using first- and second-order polynomial clutter models, while Kadi<br />
and Loupas (1995) have evaluated the use of polynomial models up to the fourth order. Torp<br />
(1997) has subsequently generalized this filtering approach by using orthogonal polynomials<br />
such as Legendre or Chebyshev polynomials † <strong>for</strong> the clutter model. In terms of its per<strong>for</strong>mance,<br />
Bjaerum et al. (2002a) have shown that, <strong>for</strong> the same filter order, the polynomial regression filter<br />
have a similar frequency response as the one <strong>for</strong> the projection-initialized IIR filter at low<br />
nominal cutoff frequencies. However, the polynomial regression filter has lesser flexibility in<br />
adjusting the width of the filter stopband since the filter response can only be changed by varying<br />
the filter order (whereas IIR filters can also vary the nominal cutoff frequency). Note that,<br />
besides the use of polynomials, other basis sets such as wavelets can also be used <strong>for</strong> the<br />
regression filter. For instance, Cloutier et al. (2003) have modeled clutter as a series of Gabor<br />
wavelets (i.e. orthogonal bases with different frequencies and Gaussian-shaped envelopes) that<br />
match the most dominant Doppler signal contents.<br />
2.2.2 Adaptive Filters<br />
Background Considerations<br />
When using non-adaptive filters <strong>for</strong> clutter suppression, it is inherently assumed that the<br />
clutter spectral contents of all the sample volumes would fall within a certain low-frequency<br />
band. Hence, the same clutter filter is often applied to the Doppler signal of all sample volumes.<br />
Nonetheless, clutter suppression would likely be more effective if the filter stopband can be<br />
selected adaptively according to the local clutter characteristics of individual sample volumes.<br />
For instance, in per<strong>for</strong>ming clutter filtering on Doppler data acquired under substantial tissue<br />
motion, the filter stopband may need to be dynamically adjusted because the Doppler clutter of<br />
some sample volumes may not be centered around zero frequency. <strong>Based</strong> on this principle, Yoo<br />
et al. (2003) have proposed the use of a filter bank approach whereby the width of the filter<br />
stopband can be changed depending on the magnitude of the mean clutter frequency. However,<br />
when a wider filter stopband is used, this clutter suppression approach may concomitantly<br />
† Legendre and Chebyshev polynomials can be obtained by orthogonalizing the polynomial set {1, n, n 2 , …}. The<br />
only difference between these two types of polynomials is their normalization weights.
Chapter 2. Clutter Suppression in <strong>Color</strong> <strong>Flow</strong> Imaging: A Hankel-SVD Approach 31<br />
(a) Clutter-Downmixing Filter<br />
Downmix<br />
clutter to f D<br />
=0<br />
Filter Response<br />
(b) Non-Adaptive Filter<br />
Clutter<br />
Spectrum<br />
Filter Response<br />
Shifted Clutter<br />
Spectrum<br />
Shifted Blood<br />
Spectrum<br />
Blood<br />
Spectrum<br />
0<br />
f D<br />
0<br />
f D<br />
Fig. 2-2. Illustration of the difference between (a) an adaptive filter with clutter downmixing<br />
and (b) a non-adaptive clutter filter. For the non-adaptive filter, a wider stopband is needed<br />
to suppress clutter, but its use causes parts of the blood spectrum to be attenuated.<br />
suppress a substantial portion of the blood echoes in the Doppler spectrum, and as a result the<br />
sensitivity of flow detection may be decreased. In order to effectively suppress clutter<br />
originating from tissue motion, it is beneficial to design filters that can adapt its stopband to the<br />
clutter spectral characteristics. This rationale is the fundamental basis of adaptive clutter filters.<br />
Clutter-Downmixing Filters<br />
In the presence of tissue motion, one intuitive way of adaptively suppressing Doppler<br />
clutter is to design and use a bandpass filter whose stopband is centered at the mean clutter<br />
frequency. To implement this approach, however, high computation power would be required<br />
because of the need to design a different bandpass filter <strong>for</strong> each sample volume. As an<br />
equivalent way of realizing the same bandpass filter operation, it is possible to first downmix the<br />
Doppler signal with the mean clutter frequency be<strong>for</strong>e carrying out a highpass filtering operation.<br />
An illustration of this clutter downmixing strategy, which was originally introduced by Thomas<br />
and Hall (1994) and Brands et al. (1995), is shown in Fig. 2-2. Mathematically, the n th sample of<br />
the downmixed Doppler signal x DM (n) can be expressed as:<br />
x<br />
− jφ<br />
( n)<br />
DM<br />
n)<br />
= x(<br />
n)<br />
e <strong>for</strong> φ(<br />
n)<br />
= 2<br />
( πf<br />
nT<br />
D(c)<br />
PRI<br />
, (2-8a)<br />
where φ(n) is the instantaneous phase of the downmixing carrier, while f D(c) denotes the mean<br />
clutter frequency that can be estimated using closed-<strong>for</strong>m methods like the lag-one autocorrelator<br />
(Kasai et al. 1985). From a vector space perspective, the downmixing operation is equivalent to<br />
applying a diagonal matrix operator Θ (of size N D ×N D ) to the Doppler signal vector x. As such,<br />
the downmixed Doppler signal vector x DM can be written as:
Chapter 2. Clutter Suppression in <strong>Color</strong> <strong>Flow</strong> Imaging: A Hankel-SVD Approach 32<br />
DM<br />
− jφ<br />
(0) − jφ<br />
(1) − jφ<br />
( ND<br />
−1)<br />
[ e , e , K,<br />
e ]<br />
x = Θx <strong>for</strong> Θ = diag<br />
. (2-8b)<br />
After the downmixing, any type of non-adaptive filter can generally be used to suppress the zerofrequency<br />
centered Doppler clutter. Consequently, the overall filter matrix of a filter with clutter<br />
downmixing is simply equal to the multiplication between the original filter matrix F and the<br />
downmixing operator Θ. As seen in Fig. 2-2, this filtering strategy is particularly useful in<br />
situations where tissue and blood scatterers move in opposite directions (e.g. when arterial<br />
distension pushes tissue towards surface while blood scatterers flow away from surface), because<br />
in this case the blood spectrum can more likely be preserved after the clutter downmixing<br />
process. Note that, as described by Bjaerum et al. (2002b), the per<strong>for</strong>mance of the downmixing<br />
filter can be further improved if instantaneous clutter frequencies are used <strong>for</strong> the phase terms in<br />
the downmixing operator. Nevertheless, to accurately estimate the instantaneous clutter<br />
frequencies, one would require the availability of multiple Doppler signal vectors that are<br />
statistically stationary.<br />
<strong>Eigen</strong>-<strong>Based</strong> Filters<br />
Another way of adaptively suppressing Doppler clutter is to directly analyze the<br />
composition of the Doppler signal and remove a composition subset that corresponds to clutter.<br />
This filtering strategy is essentially the same as using a regression filter whose clutter model<br />
consists of a subset of components seen in the Doppler signal composition. In terms of its<br />
implementation, the signal analysis can be effectively carried out by decomposing the Doppler<br />
data into a series of adaptable, orthogonal basis functions (as opposed to fixed ones like the<br />
Fourier expansion). As discussed in algebra textbooks (e.g. see Moon and Stirling 2000, Sec.<br />
6.8), such decomposition is often referred to as the Karhunen-Loeve (KL) expansion or principal<br />
component analysis, and it can be computed through an eigen-decomposition of the Doppler<br />
signal’s correlation matrix. For a Doppler ensemble size N D , the KL expansion is given by:<br />
ND<br />
⎧λk<br />
, k = l<br />
x = ∑γ ke<br />
k<br />
<strong>for</strong> E{ γ<br />
kγ<br />
l}<br />
= ⎨ . (2-9a)<br />
k = 1<br />
⎩ 0 , k ≠ l<br />
where γ k is the k th expansion coefficient, while λ k and e k are the eigenvalue and eigenvector <strong>for</strong><br />
the k th orthogonal basis function. Note that λ k and e k are related to the Doppler signal as follows:<br />
ND<br />
∑<br />
k = 1<br />
λ R , (2-9b)<br />
H<br />
H<br />
ke ke<br />
k<br />
= E{ xx } =<br />
x
Chapter 2. Clutter Suppression in <strong>Color</strong> <strong>Flow</strong> Imaging: A Hankel-SVD Approach 33<br />
in which R x =E{xx H } refers to the statistical correlation matrix of the Doppler signal. In practice,<br />
this matrix can be estimated via ensemble averaging as follows:<br />
M<br />
1<br />
R ≈ x x , (2-9c)<br />
x<br />
∑<br />
H<br />
m m<br />
M m=<br />
1<br />
where M is the number of Doppler signal vectors (or snapshots) that are statistically stationary<br />
and x m indicates the Doppler signal vector <strong>for</strong> the m th realization.<br />
Since clutter often has a much higher strength than blood echoes and white noise, the<br />
signal decomposition given in (2-9a) would have high-energy basis functions (or principal<br />
components) that correspond to clutter. Hence, as pointed out by Bjaerum et al. (2002b),<br />
Doppler clutter can be suppressed by using the basis vectors of these high-energy components as<br />
the clutter model <strong>for</strong> a regression filter. In particular, when K of the N D basis functions<br />
correspond to clutter (i.e. with a clutter eigen-space dimension equal to K), the resulting clutter<br />
model and the clutter subspace matrix <strong>for</strong> an eigen-based regression filter can be expressed as:<br />
⎡ | | | ⎤<br />
K<br />
cˆ = ⇔ =<br />
⎢<br />
⎥<br />
∑ χ<br />
ke<br />
k<br />
C<br />
⎢<br />
e1<br />
e<br />
2<br />
L e<br />
K<br />
⎥<br />
, (2-10)<br />
k = 1<br />
⎢⎣<br />
| | | ⎥⎦<br />
and the filter matrix would carry the same <strong>for</strong>m shown in (2-6b). Similar <strong>for</strong>ms of this eigenbased<br />
filter have been reported in studies on swept-scan-based flow imaging (Kruse and Ferrara<br />
2002) as well as strain-flow imaging (Kargel et al. 2003). Note that, as described by Ledoux et<br />
al. (1997), the same filtering strategy can also be per<strong>for</strong>med by applying singular value<br />
decomposition (SVD) to a multi-snapshot data matrix constructed from stacking together a<br />
number of stationary Doppler signal vectors. As well, as reported by Gallippi and Trahey<br />
(2002), the Doppler spectral analysis can be carried out using another approach called<br />
independent component analysis that decomposes a signal into a series of statistically<br />
independent basis functions (as opposed to orthogonal ones). This latter approach has been<br />
applied to suppress Doppler clutter in acoustic radiation <strong>for</strong>ce imaging (Gallippi et al. 2003).<br />
2.2.3 Comparison of Clutter Filters<br />
Table 2-1 summarizes and compares the main features of the four types of clutter filters<br />
described in this section. From this table, it can be seen that the two non-adaptive filters<br />
(projection-initialized IIR and polynomial regression) are simpler to implement than adaptive
Chapter 2. Clutter Suppression in <strong>Color</strong> <strong>Flow</strong> Imaging: A Hankel-SVD Approach 34<br />
Type<br />
IIR + Proj.<br />
Initialization<br />
Table 2-1. Comparison of Existing Clutter Filter Designs<br />
Stopband<br />
Adjustment<br />
Through changes<br />
in filter order and<br />
nominal cutoff<br />
Advantage<br />
Simple to carry out once<br />
filter coefficients are found<br />
Limitation<br />
Does not adapt to Doppler<br />
signal; efficacy depends on<br />
choice of filter parameters<br />
Polynomial<br />
Regression<br />
By choice of<br />
clutter polynomial<br />
order<br />
Same as above<br />
Same as above<br />
Clutter<br />
Downmixing<br />
From changes to<br />
stopband of nonadaptive<br />
filter<br />
Improves highpass filtering<br />
by first downmixing clutter<br />
to zero frequency<br />
Efficacy depends on<br />
choice of non-adaptive<br />
filter and accuracy of<br />
downmixing matrix<br />
<strong>Eigen</strong>-<br />
Regression<br />
Via selection of<br />
clutter eigen-space<br />
dimension<br />
Adapts filter matrix to<br />
Doppler signal contents via<br />
KL expansion<br />
Needs multiple snapshots<br />
to find correlation matrix;<br />
efficacy depends on choice<br />
of clutter dimension<br />
filters because their filter matrix remains unchanged <strong>for</strong> a given set of filter parameters.<br />
However, the primary shortcoming of these non-adaptive filters is that they are not adapted to the<br />
Doppler signal contents. Because of this limitation, they inherently need a wider stopband to<br />
suppress Doppler clutter that has wideband characteristics and ones that are shifted away from<br />
zero frequency. On another note, it is worth pointing out that amongst the two non-adaptive<br />
filters, the projection-initialized IIR filter seems to have more flexibility in defining the stopband<br />
since it can modify the filter response via changes in both the filter order and the nominal cutoff<br />
frequency.<br />
In contrast to the two non-adaptive filters, the clutter-downmixing filter can be perceived<br />
as a partially adaptive filter. Specifically, its filter matrix is a joint product between a fixed<br />
matrix operator that makes use of non-adaptive filtering principles as well as an adaptive matrix<br />
operator that is intended to downmix the Doppler clutter to zero frequency. The advantage of<br />
this filtering approach is that the highpass filtering operation can likely be improved since the<br />
downmixed Doppler clutter is supposedly centered at zero frequency. Nevertheless, because of<br />
the hybrid nature of its filter matrix, the clutter-downmixing filter’s efficacy is inherently<br />
dependent on two factors: 1) the choice of the non-adaptive filter used <strong>for</strong> the fixed matrix
Chapter 2. Clutter Suppression in <strong>Color</strong> <strong>Flow</strong> Imaging: A Hankel-SVD Approach 35<br />
operator, and 2) the accuracy of the downmixing matrix operator found by estimating the mean<br />
or instantaneous clutter frequencies of the Doppler data.<br />
Amongst the four classes of filters considered, the eigen-regression filter seems to have<br />
the best adaptability to the Doppler signal contents. In particular, the filter matrix of this filter is<br />
defined adaptively according to the most principal basis vectors in the Doppler signal’s KL<br />
expansion. In turn, the eigen-regression filter can more likely preserve non-clutter components<br />
of the Doppler signal unless they share the same principal basis vectors with clutter components.<br />
However, the <strong>for</strong>mulation <strong>for</strong> this filter inherently requires the availability of Doppler data from<br />
multiple sample volumes with similar signal characteristics (i.e. statistically stationarity). As<br />
such, this filter may be challenging to implement in some vascular imaging studies where it is<br />
difficult to segment out regions with similar Doppler data statistics. Another limitation of the<br />
eigen-regression filter is that its efficacy is rather sensitive to the choice of the clutter eigenspace<br />
dimension owing to the filter’s adaptive nature. Hence, it is necessary to develop an<br />
effective algorithm <strong>for</strong> this filter to select the clutter eigen-space dimension.<br />
2.3 The Hankel-SVD Filter<br />
2.3.1 Background Considerations<br />
Design Motivations<br />
Despite its adaptability to the Doppler signal contents, the eigen-based filtering strategy<br />
embodied in (2-9) and (2-10) is often faced with two types of challenges in practice (as noted in<br />
the previous section). First, consistent estimation of Doppler correlation matrix R x seen in (2-9c)<br />
generally requires multiple Doppler signal vectors that are statistically stationary. In the reported<br />
eigen-based filter designs, this estimation is often per<strong>for</strong>med by using the Doppler signal vectors<br />
along the same beam line (with the inherent assumption that the clutter statistics are stationary<br />
over the depth of view). Such an approach is generally effective in microvasculature flow<br />
studies where the imaging depth is only of a few millimeters, but it might not be valid <strong>for</strong> general<br />
vascular studies where the imaging depth can be several centimeters. Second, precise<br />
determination of the clutter eigen-space dimension K as seen in (2-10) is essential to the efficacy<br />
of the eigen-based filter. In most of the previously reported eigen-based filter designs, this<br />
parameter was manually selected by the user during analysis according to some prior knowledge
Chapter 2. Clutter Suppression in <strong>Color</strong> <strong>Flow</strong> Imaging: A Hankel-SVD Approach 36<br />
of the clutter spectral characteristics. Such a manual way of specifying the clutter dimension<br />
may reduce the filtering per<strong>for</strong>mance in settings where the clutter spectral characteristics<br />
generally vary over time. To avoid these practical challenges, a new eigen-based filtering<br />
strategy will be developed in this section by exploiting the eigen-space properties of the Hankel<br />
matrix (which has constant entries along its reverse diagonals).<br />
Fundamental Principles<br />
The new filtering approach, which we refer to as the Hankel-SVD filter, is derived from<br />
insights † drawn upon a principal Hankel component analysis framework that is often used in<br />
time series analysis (van der Veen et al. 1993). Like the KL expansion, this new method<br />
attempts to decompose the Doppler signal as a sum of adaptable, orthogonal basis functions.<br />
However, instead of relying on multiple snapshots of Doppler signal to estimate the basis<br />
functions, the Hankel-SVD approach computes the basis functions through SVD analysis of the<br />
Hankel data matrix created from the Doppler signal of individual sample volumes. It should be<br />
noted that, because of the Hankel matrix’s inherent structure, the resulting Doppler signal<br />
composition would only have P orthogonal bases (with P less than N D ). As such, the signal<br />
synthesis equation <strong>for</strong> the Hankel-SVD approach can be expressed as:<br />
P<br />
x = ∑γ φ <strong>for</strong> γ ⊥ γ , (2-11)<br />
k = 1<br />
k<br />
k<br />
where γ k and ϕ k are respectively the k th expansion coefficient and orthonormal basis vector in the<br />
Hankel component approximation. This <strong>for</strong>mulation should be distinguished from the SVD<br />
analysis framework reported by Ledoux et al. (1997), who used multiple Doppler signal vectors<br />
to construct the data matrix.<br />
2.3.2 Theoretical Formulation<br />
Construction of Hankel Data Matrix<br />
An overview of the new eigen-based filtering strategy is depicted in Fig. 2-3. In terms of<br />
its details, the Hankel-SVD filter begins with the <strong>for</strong>mation of a Hankel data matrix by dividing<br />
the Doppler signal into partially overlapping segments and rearranging them into an array. For a<br />
k<br />
l<br />
† This new framework is also motivated by our investigations on the Matrix Pencil estimation framework (to be<br />
presented in Chapter 3).
Chapter 2. Clutter Suppression in <strong>Color</strong> <strong>Flow</strong> Imaging: A Hankel-SVD Approach 37<br />
Doppler<br />
<strong>Signal</strong><br />
x<br />
Create Hankel<br />
Matrix<br />
A =<br />
0 1<br />
2<br />
3<br />
1<br />
2<br />
3<br />
4<br />
2<br />
3<br />
3<br />
4<br />
4<br />
5<br />
5<br />
6<br />
4 5 6<br />
7<br />
Per<strong>for</strong>m SVD<br />
P<br />
A = Σσ k<br />
u k<br />
v<br />
H<br />
k<br />
1<br />
Find Principal Basis<br />
Functions<br />
γ k ϕ k (n) =<br />
ΣΣa k (i,j) δ n,i+j–1<br />
ΣΣ δ n,i+j–1<br />
Filter Output<br />
y<br />
Reconstruct<br />
Filtered <strong>Signal</strong><br />
P<br />
y = Σ γ k ϕ k<br />
K c +1<br />
Estimate Clutter <strong>Eigen</strong>-Space<br />
Dimension<br />
Is |f D,1 | > f thr(clut) ?<br />
No<br />
Yes<br />
K c = 0<br />
K c =argmax{|f D,k – f D,1 |
Chapter 2. Clutter Suppression in <strong>Color</strong> <strong>Flow</strong> Imaging: A Hankel-SVD Approach 38<br />
Doppler signal vector x<br />
0<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
Hankel matrix A<br />
0 1 2 3<br />
1 2 3 4<br />
2<br />
3<br />
3<br />
4<br />
4<br />
5<br />
5<br />
6<br />
4 5 6 7<br />
Fig. 2-4. Illustration of how the Hankel matrix A is constructed <strong>for</strong> the case where the<br />
Doppler ensemble size is eight (N D =8) and the dimension parameter is four (P=4).<br />
where A k is the k th rank-one Hankel component in the decomposition, while σ k , u k , and v k are<br />
correspondingly the singular value, left singular vector (with dimension N D –P+1), and right<br />
singular vector (with dimension P) of A k . As discussed in algebra textbooks (e.g. Moon and<br />
Stirling 2000, Sec. 7.1), the P singular values in (2-13) are ordered from largest to smallest by<br />
definition, and hence A k can be considered as the k th -order principal Hankel component. Note<br />
that, by using these principal Hankel components, it is possible to reconstruct the orthogonal<br />
basis functions γ k ϕ k as seen in (2-11). One particular way to per<strong>for</strong>m this reconstruction process,<br />
as originally described by Poon et al. (1993), is to sum and average the matrix elements along the<br />
reverse diagonals of A k . <strong>Based</strong> on this notion, the n th sample of γ k ϕ k (n) can be mathematically<br />
expressed as:<br />
N −P<br />
P−1<br />
D<br />
∑∑<br />
n=<br />
i+<br />
j<br />
i=<br />
0 j=<br />
0<br />
=<br />
N −P<br />
P−1<br />
D<br />
∑∑<br />
i=<br />
0<br />
δ<br />
j=<br />
0<br />
a ( i,<br />
j)<br />
k<br />
γ<br />
kϕk<br />
( n)<br />
, (2-14)<br />
δ<br />
where a k (i, j) is the entry of A k at row i and column j and δ n=i+j is a Kronecker delta function that<br />
equals to one if n = i+j and equals to zero otherwise.<br />
Estimation of Clutter <strong>Eigen</strong>-Space Dimension<br />
Since the aim of the Hankel-SVD filter is to suppress clutter in the Doppler signal, it is<br />
necessary to determine whether a principal Hankel component A k is part of the clutter eigenspace.<br />
In general, there are two types of approaches to carry out this analysis. First, given that<br />
clutter often has higher energy than blood echoes and white noise, a principal Hankel component<br />
can be considered as being part of the clutter eigen-space if its singular value magnitude σ k is<br />
n=<br />
i+<br />
j
Chapter 2. Clutter Suppression in <strong>Color</strong> <strong>Flow</strong> Imaging: A Hankel-SVD Approach 39<br />
1. Test clutter presence with f thr(clut)<br />
2. Identify clutter dimension with Δf thr<br />
Component Mean Freq. (f D,k<br />
)<br />
f D,1<br />
0<br />
+f thr(clut)<br />
k<br />
–f thr(clut)<br />
Component Mean Freq. (f D,k<br />
)<br />
f D,1<br />
0<br />
Clutter Band<br />
k<br />
Δf thr<br />
Fig. 2-5. Steps used by the Hankel-SVD filter to estimate the clutter eigen-space dimension.<br />
First (left figure), clutter presence is examined by testing whether the most dominant<br />
component has a mean frequency (marked as black ⊗) less than a threshold f thr(clut) . After that<br />
(right figure), clutter dimension is determined by finding largest component order with mean<br />
frequency estimate inside the clutter band Δf thr (in this case, the clutter dimension is three).<br />
larger than a given value. Examples of clutter eigen-space estimation algorithms using this<br />
analysis approach can be found in a few previously reported eigen-filter designs (Kruse and<br />
Ferrara 2002, Lovstakken et al. 2006). Alternatively, since clutter generally consists of lowfrequency<br />
contents, it is possible to identify a clutter eigen-space component based on the<br />
frequency contents of each principal basis vector ϕ k . This latter approach is used in the Hankel-<br />
SVD filter to determine the clutter eigen-space dimension, since it can be implemented in a way<br />
that gives filtering characteristics similar to the stopband of a conventional bandpass filter.<br />
The clutter eigen-space analysis algorithm used by the Hankel-SVD filter is illustrated in<br />
Fig. 2-5. This algorithm is fundamentally based on two assumptions: 1) Doppler clutter is<br />
contained in the more dominant Hankel components (i.e. ones with larger singular values); 2) the<br />
blood flow component of the Doppler signal is contained in the Hankel components with high<br />
Doppler frequencies. In terms of its details, the analysis begins with an assessment of whether<br />
clutter is present in the Doppler signal by finding whether the most dominant Hankel component<br />
has a mean frequency f D,1 that is less than a spectral threshold f thr(clut) . If clutter is considered to<br />
be present, then the mean frequency of the other principal basis functions are estimated. Note<br />
that, by using the well-known single-lag autocorrelator (Kasai et al. 1985), the mean Doppler<br />
frequency f D,k corresponding to each principal basis function ϕ k can be found to be equal to:
Chapter 2. Clutter Suppression in <strong>Color</strong> <strong>Flow</strong> Imaging: A Hankel-SVD Approach 40<br />
f<br />
D, k<br />
ND<br />
⎡<br />
⎤<br />
= ⎢ ∑ − 1<br />
1<br />
*<br />
arg ϕ<br />
k<br />
( n)<br />
ϕ<br />
k<br />
( n −1)<br />
⎥ . (2-15)<br />
2πTPRI<br />
⎣ n=<br />
1<br />
⎦<br />
From these estimates, the clutter eigen-space dimension is found by searching <strong>for</strong> the largest<br />
component order whose mean frequency falls within a clutter band Δf thr that is centered at f D,1 (as<br />
shown in Fig. 2-5). Consequently, the estimated clutter dimension K c can be expressed as:<br />
K<br />
c<br />
⎪<br />
⎧<br />
= ⎨arg max<br />
⎪⎩ k<br />
{ f − f < Δf<br />
/ 2}<br />
D, k<br />
0<br />
D,1<br />
thr<br />
,<strong>for</strong><br />
, <strong>for</strong><br />
f<br />
f<br />
D,1<br />
D,1<br />
><br />
<<br />
f<br />
f<br />
thr(clut)<br />
thr(clut)<br />
. (2-16)<br />
Once the clutter dimension has been estimated, the filtered Doppler signal can be found as<br />
follows by summing the principal basis functions beyond the estimated clutter dimension:<br />
P<br />
∑<br />
y = γ φ . (2-17)<br />
k<br />
k = KC<br />
+ 1<br />
It is worth noting that the average power of the filtered Doppler signal can be estimated from the<br />
singular values of the principal Hankel components beyond the clutter dimension. In particular,<br />
since the squared sum of singular values is equal to the Frobenius matrix norm (Moon and<br />
Stirling 2000, Sec. 7.1), the post-filter Doppler power can be estimated to equal to:<br />
P<br />
2<br />
k<br />
∑<br />
k = K + 1 P(<br />
N<br />
D<br />
− P +<br />
k<br />
σ<br />
ρ y<br />
=<br />
. (2-18)<br />
1<br />
C<br />
)<br />
where the normalization factor in the denominator stems from the fact that each Hankel<br />
component A k has P(N D –P+1) entries.<br />
Computational Considerations<br />
As can be expected from its <strong>for</strong>mulation, the computation load of the Hankel-SVD filter<br />
is rather high comparing to non-adaptive filters like the IIR filter. In particular, this new filtering<br />
strategy requires the computation of an SVD on the Hankel matrix <strong>for</strong>med from each Doppler<br />
signal, and after the SVD is computed, a series of vector operations is needed to reconstruct the<br />
principal Hankel components. To estimate the overall computation burden of the new filter, an<br />
analysis can be per<strong>for</strong>med to determine the number of floating point operations † (flops) required<br />
during various stages of the filter. Such an analysis is provided in Table 2-2. As can be seen,<br />
applying the Hankel-SVD filter to each Doppler signal would need at least on the order of P 3<br />
† In this study, both multiplication and addition are considered as floating point operations.
Chapter 2. Clutter Suppression in <strong>Color</strong> <strong>Flow</strong> Imaging: A Hankel-SVD Approach 41<br />
Table 2-2. Summary on the Flops Needed in the Hankel-SVD Filter<br />
Step Flops Rationale<br />
1. Form Hankel matrix<br />
of size (N D –P+1)×P<br />
2. Compute SVD of the<br />
Hankel matrix<br />
3. Reconstruct the<br />
principal basis<br />
functions<br />
0<br />
O[P 3 ]<br />
O[P 3 ]<br />
• No computations required to <strong>for</strong>m the Hankel data<br />
matrix<br />
• According to Golub and Van Loan (1996, Sec. 8.3),<br />
SVD of an (N D –P+1)×P matrix needs<br />
O[(N D –P+1)P 2 ] flops<br />
• For P=N D /2, O[(N D –P+1)P 2 ] ≈ O[P 3 ]<br />
• O[(N D –P+1)P] flops needed to reconstruct ϕ k and<br />
A k from σ k u k v k<br />
H<br />
• P principal components to reconstruct, so<br />
O[(N D –P+1)P 2 ] flops needed in total<br />
• For P=N D /2, O[(N D –P+1)P 2 ] ≈ O[P 3 ]<br />
4. Estimate component<br />
mean frequencies<br />
5. Estimate clutter<br />
eigen-space dimension<br />
6. Reconstruct filter<br />
output<br />
O[P 2 ]<br />
0<br />
O[N D ]<br />
• O[N D ] flops needed to compute the lag-one<br />
autocorrelation estimate<br />
• With P components to estimate, O[N D P] flops<br />
needed in total<br />
• For P=N D /2, O[N D P] ≈ O[P 2 ]<br />
• No computations required to estimate clutter<br />
dimension<br />
• O[N D ] flops required to sum all principal<br />
components beyond the clutter dimension<br />
flops (when P is set to N D /2). This computational demand is one exponential order higher than<br />
that required <strong>for</strong> typical non-adaptive filters (by inspection of (2-2), such a filter only needs on<br />
the order of N 2 D flops). Nevertheless, since N D is often kept below 20 samples in color flow<br />
imaging, the additional burden needed by the Hankel-SVD filter is less substantial. As well,<br />
with the persymmetric structure of the Hankel data matrix, the SVD can be computed with lesser<br />
flops via the use of more efficient algorithms (Badeau et al. 2004).<br />
2.4 Simulation Method<br />
2.4.1 <strong>Signal</strong> Synthesis Model<br />
General Principles<br />
To facilitate analysis of the Hankel-SVD filter’s per<strong>for</strong>mance, a signal synthesis model<br />
was first developed to generate Doppler signals with various Doppler spectral characteristics. In
Chapter 2. Clutter Suppression in <strong>Color</strong> <strong>Flow</strong> Imaging: A Hankel-SVD Approach 42<br />
this synthesis model, the clutter and blood components of the Doppler signal were separately<br />
generated using two different approaches. For the clutter components, a phase modulation<br />
model was used to synthesize ensembles with non-stationary clutter spectral characteristics. As<br />
first introduced by Heimdal and Torp (1997), the physical notion behind this synthesis approach<br />
is that, due to the contractile nature of muscles, body tissues tend to move in a cyclic pattern and<br />
in turn give rise to Doppler clutter with phase modulation features. In contrast, <strong>for</strong> the blood<br />
flow components of the Doppler signal, they were generated by feeding complex Gaussian noise<br />
through a linear filter whose impulse response corresponds to the Doppler signal <strong>for</strong> a single<br />
moving scatterer. As originally described by Kristoffersen and Angelsen (1988), the physical<br />
rationale behind this filter-based synthesis approach is that, due to the random nature of blood<br />
scatterer distribution, the scattering strength at each Doppler sampling instant is essentially a<br />
complex Gaussian random variable. As such, the overall Doppler signal from various blood<br />
scatterers can be modeled as a superposition of various randomly time-lagged and amplitudescaled<br />
replica of the single-scatterer echo template (i.e. similar to a convolution or linear filtering<br />
operation). Note that, as reviewed by Mo and Cobbold (1992), this blood signal synthesis model<br />
has the advantage of allowing the user to directly control the Doppler spectral characteristics of<br />
blood scatterers, and thus it is useful <strong>for</strong> testing various Doppler signal processing strategies.<br />
Synthesis Procedure<br />
Fig. 2-6 shows a system-level schematic of the procedure used in the synthesis model. In<br />
general, the entire synthesis procedure can be divided into three main parts. In the first part, the<br />
clutter component of Doppler signal was generated. This process begins by constructing a<br />
clutter vibration pattern to represent the instantaneous phase of the Doppler clutter. For our<br />
study, the vibration pattern was modeled as a single-tone wave<strong>for</strong>m whose frequency and peak<br />
amplitude were set based on the specified clutter parameters (to be described in the next<br />
subsection). The clutter signal template was then obtained by setting the vibration pattern as the<br />
phase argument of a complex exponential operator, and correspondingly the n th sample of this<br />
template can be expressed as follows:<br />
h ( n)<br />
c<br />
jφ<br />
sin(2πf<br />
nT<br />
)<br />
c,max vib PRI<br />
= e<br />
, (2-19)<br />
where φ c,max is the peak phase deviation in the phase-modulated signal and f vib is the vibration<br />
frequency of the clutter vibration pattern. To model the random magnitude and phase of the
Chapter 2. Clutter Suppression in <strong>Color</strong> <strong>Flow</strong> Imaging: A Hankel-SVD Approach 43<br />
Clutter<br />
Vibration<br />
Pattern<br />
Complex<br />
Exponential<br />
Operator<br />
Clutter<br />
Strength<br />
Σ<br />
Clutter<br />
Doppler<br />
<strong>Signal</strong><br />
Gaussian<br />
Noise<br />
Blood<br />
Scatterer<br />
Distribution<br />
Single-Scatterer<br />
Echo Template<br />
Noise<br />
Strength<br />
Σ<br />
Clutter + Blood<br />
Doppler<br />
<strong>Signal</strong><br />
Blood<br />
Strength<br />
Fig. 2-6. System-level schematic of the Doppler signal synthesis approach. In this study, the<br />
single-scatterer blood echo template carried the <strong>for</strong>m of a Gaussian-shaped complex sinusoid,<br />
while the clutter vibration pattern was a single-tone wave<strong>for</strong>m.<br />
tissue scatterer distribution, the clutter signal template was subsequently multiplied with a<br />
complex Gaussian random number.<br />
In the second part of the synthesis process, the blood component of the Doppler signal<br />
was generated. This process first involves the creation of an echo template to model the Doppler<br />
signal originating from a single moving scatterer. The echo template used in this study was<br />
defined as a Gaussian-enveloped complex sinusoid whose modulating frequency corresponds to<br />
the mean Doppler frequency calculated from the specified blood flow parameters (to be<br />
described in the next subsection). For a given modulating frequency f D(b) and a temporal width<br />
parameter T dur , the n th sample of the blood echo template can be expressed as follows:<br />
h<br />
b<br />
( n)<br />
2<br />
−(<br />
nT<br />
2<br />
PRI / Tdur<br />
) j πfD(b)<br />
nTPRI<br />
= e e . (2-20)<br />
To account <strong>for</strong> the finite transit time of blood scatterers (as recently reviewed by Yu et al.<br />
(2006)), the temporal width of the blood echo template was defined as a function of blood flow<br />
velocity. From the single-scatterer echo template, the blood signal <strong>for</strong> multiple scatterers was<br />
then synthesized by convolving complex Gaussian random samples with the template. Note that,<br />
during the simulations, the transient samples in the convolution were discarded to maintain
Chapter 2. Clutter Suppression in <strong>Color</strong> <strong>Flow</strong> Imaging: A Hankel-SVD Approach 44<br />
statistical uni<strong>for</strong>mity in the synthesized data. As well, multiple realizations of the blood signal<br />
were generated at once by convolving a large number of random samples with the singlescatterer<br />
echo template.<br />
In the last part of the synthesis process, the blood and clutter components generated from<br />
the previous parts were combined to <strong>for</strong>m the Doppler signal. To carry out this process, the<br />
amplitudes of the blood and clutter components were first scaled according to the specified<br />
clutter and blood scattering strengths. After that, the scaled components were then summed<br />
together along with white noise samples that were generated separately. For analysis purpose,<br />
the clutter-only Doppler signal (i.e. the signal <strong>for</strong> a sample volume located in the tissue region)<br />
was also obtained by only summing the scaled clutter components and white noise samples.<br />
Note that the n th sample of the synthesized Doppler signal and its clutter-only counterpart can be<br />
expressed as follows:<br />
[ g h ( n)<br />
] + κ [ g ( n)<br />
∗ h ( n)<br />
] κ w(<br />
n)<br />
x( n)<br />
= κ<br />
c c c<br />
b b b<br />
+<br />
w<br />
, (2-21a)<br />
[ g h ( n)<br />
] κ w(<br />
n)<br />
xc<br />
( n)<br />
= κ<br />
c c c<br />
+<br />
w<br />
, (2-21b)<br />
where g c and g b (n) are respectively the complex Gaussian random samples used to model the<br />
tissue and blood scatterer distributions; also, κ c , κ b , and κ w respectively denote the amplitude<br />
scaling coefficients <strong>for</strong> clutter, blood echoes, and white noise. Prior to using the signals<br />
synthesized from (2-21a) and (2-21b) <strong>for</strong> filter per<strong>for</strong>mance analysis, they were quantized to a<br />
certain number of bits as set <strong>for</strong>th by the specified dynamic range, and they were divided into<br />
non-overlapping segments to yield multiple signal realizations.<br />
Model Parameters<br />
In the simulation model, four main classes of parameters can be specified during the<br />
synthesis process so that Doppler signals with various spectral characteristics can be generated.<br />
The first class of parameters is associated with the ultrasound machine and the synthesis process.<br />
These parameters include acoustic speed (c o ), pulse carrier frequency (f o ), pulse repetition<br />
interval (T PRI ), F-number (i.e. ratio of the aperture width to the focal length, F/W), Doppler<br />
ensemble size (N D ), dynamic range of data sampling, and number of realizations to be<br />
synthesized. As already discussed in Section 1.4.2, most of these system parameters have a<br />
direct influence on the velocity aliasing limit and the velocity resolution of the synthesized<br />
Doppler signals.
Chapter 2. Clutter Suppression in <strong>Color</strong> <strong>Flow</strong> Imaging: A Hankel-SVD Approach 45<br />
The second class of parameters is related to the generation of the clutter signal. In<br />
particular, the two clutter parameters of interest are clutter vibration frequency (f vib ) and<br />
maximum tissue velocity (v c,max ). As its name implies, the clutter vibration frequency controls<br />
the rate of change of the clutter signal’s instantaneous phase. This parameter, along with the<br />
maximum tissue velocity, also influences the peak phase deviation in the phase-modulated<br />
clutter signal. Following a similar derivation by Heimdal and Torp (1997), it can be shown that<br />
the two clutter parameters influence the peak clutter phase deviation (φ c,max ) according to the<br />
following relationship:<br />
2 f<br />
ovc,max<br />
φ<br />
c,max<br />
= . (2-22)<br />
c f<br />
Note that, as discussed in various communications textbooks (e.g. see Haykin 2001, Sec. 2.7),<br />
the phase-modulated signal can have time-varying spectral components anywhere over spectral<br />
range ±2f vib (1+φ c,max ).<br />
The third class of parameters is concerned with the synthesis of the blood signal.<br />
Specifically, the two blood parameters of interests are mean flow velocity (v b ) and beam-flow<br />
angle (θ). According to the Doppler equation given in (1-1), these two parameters directly<br />
govern the Doppler modulation frequency f D(b) in the single-scatterer blood echo template. Also,<br />
<strong>for</strong> large beam-flow angles, it can be shown that the two parameters affect the bandwidth of the<br />
blood echo template as follows due to transit time broadening (Yu et al. 2006):<br />
2Wv<br />
o<br />
f<br />
vib<br />
sinθ<br />
b o<br />
β<br />
b<br />
= . (2-23a)<br />
Fco<br />
Assuming that this value represents the –20 dB spectral broadening bandwidth, the temporal<br />
width parameter given in (2-20) can then be shown to be equal to the following based on the<br />
Fourier trans<strong>for</strong>m of Gaussian functions and some trivial derivations:<br />
2 ln10<br />
Fco<br />
T<br />
width<br />
≈ = 0.483<br />
. (2-23b)<br />
πβ Wv f sinθ<br />
b<br />
The last class of parameters is related to the relative strengths of the clutter, blood echoes,<br />
and white noise. In the simulation model, these relative strengths are characterized according to<br />
the clutter-to-blood signal ratio (CBR) and the blood-signal-to-noise ratio (BSNR), which in turn<br />
are defined as follows:<br />
b<br />
o
Chapter 2. Clutter Suppression in <strong>Color</strong> <strong>Flow</strong> Imaging: A Hankel-SVD Approach 46<br />
Table 2-3. Simulation Parameters Used in the Clutter Filter Study<br />
Parameter Arterial flow Low-velocity flow<br />
Fixed Machine and Synthesis Parameters<br />
Acoustic speed, c o<br />
1540 m/s<br />
F-number, F/W 4<br />
Doppler ensemble size, N D 10, 20<br />
Dynamic range<br />
14 bits<br />
Number of ensembles per dataset 10000 (repeated <strong>for</strong> every increment in v b )<br />
<strong>Signal</strong> Strength Parameters<br />
Clutter-to-blood signal ratio<br />
30 dB<br />
Blood-signal-to-noise ratio<br />
10 dB<br />
Average noise strength, κ n<br />
10 dB<br />
Blood Parameters<br />
Beam-flow angle, θ 60°<br />
<strong>Flow</strong> velocity, v b<br />
Varying from 0 to aliasing velocity<br />
Manipulated Machine Parameters<br />
Pulse carrier frequency, f o 5 MHz 2 MHz<br />
Pulse repetition interval, T PRI 0.4 ms 2.0 ms<br />
Clutter Parameters<br />
Maximum tissue velocity, v c,max 2 mm/s 20 mm/s<br />
Vibration frequency, f vib 5 Hz 5 Hz<br />
Other Responding Parameters<br />
Aliasing velocity, v alias 38.5 cm/s 19.2 cm/s<br />
Clutter phase modulation index, φ c,max 2.6 radians 10.4 radians<br />
Peak clutter frequency 36 Hz 114 Hz<br />
( κ κ )<br />
CBR = 20log 10<br />
/ , (2-24a)<br />
c<br />
b<br />
( κ κ )<br />
BSNR = 20log 10<br />
/ , (2-24b)<br />
where κ c , κ b and κ n are the amplitude coefficients as seen in (2-21a) (i.e. the signal equation <strong>for</strong><br />
the synthesized Doppler signal). Note that, in this study, the white noise scaling coefficient κ n<br />
was set such that the synthesized signals were contained within the system’s dynamic range.<br />
2.4.2 Method of Analysis<br />
General Overview<br />
To assess the efficacy of the Hankel-SVD filter, various sets of Doppler signals and their<br />
clutter-only components (each with 10000 realizations) were generated using the above<br />
described synthesis model and the parameters listed in Table 2-3. The synthesized data were<br />
intended to model the received Doppler signal in a normal arterial flow scenario with minor<br />
tissue motion as well as a low-velocity flow scenario with significant tissue motion and longer<br />
b<br />
w
Chapter 2. Clutter Suppression in <strong>Color</strong> <strong>Flow</strong> Imaging: A Hankel-SVD Approach 47<br />
ensemble periods. As an initial assessment, particular signal realizations were selected from the<br />
synthesized data to analyze the characteristics of the principal Hankel components. To more<br />
quantitatively assess the filter’s per<strong>for</strong>mance, the Hankel-SVD filter was applied to all the<br />
synthesized Doppler signals and their resulting post-filter power was computed. As a benchmark<br />
<strong>for</strong> the post-filter power estimates, the same procedure was repeated with the clutter-only signals.<br />
Note that during the analysis, the dimension parameter P of the Hankel-SVD filter was set as<br />
N D /2, while the two thresholds f thr(clut) and Δf thr were respectively set to 100 Hz and 50 Hz.<br />
Per<strong>for</strong>mance Measure<br />
In this study, the post-filter clutter-to-blood signal ratio (CBR) was analyzed at various<br />
blood velocities to quantitatively examine the efficacy of the Hankel-SVD filter. This quantity<br />
was computed by finding the square-rooted ratio between the average post-filter power of<br />
clutter-only signals and that of Doppler signals corresponding to a particular blood velocity. If a<br />
filter can effectively suppress clutter, the post-filter CBR should be less than one in linear scale<br />
(or be negative in dB scale) because blood echoes should dominate the filtered Doppler signal.<br />
To facilitate comparative assessment, the above filter analysis was also carried out with<br />
an IIR-based (with projection-initialization) clutter-downmixing filter. For the arterial flow<br />
scenario, the IIR filter used was either a third-order (<strong>for</strong> N D =10) or fourth-order (<strong>for</strong> N D =20)<br />
Chebychev filter with 100 Hz cutoff; as <strong>for</strong> the low-velocity flow scenario, filter orders of five<br />
(<strong>for</strong> N D =10) and eight (<strong>for</strong> N D =20) were used. Note that these filter orders were chosen because<br />
their stopband width is similar to the cutoff bandwidth used by the Hankel-SVD filter. It is also<br />
worth pointing out that the eigen-regression filter was not examined in these simulations because<br />
our signal synthesis model does not generate statistically-stationary Doppler ensembles <strong>for</strong><br />
multiple sample volumes (as needed by the eigen-filter to estimate the correlation matrix).<br />
Analysis of this filter will be considered in the experimental studies presented in Chapter 4.<br />
2.5 Simulation Results<br />
2.5.1 Characteristics of Principal Hankel Components<br />
Arterial <strong>Flow</strong> Scenario<br />
To study the characteristics of principal Hankel components, one particular realization of<br />
Doppler signal (with an ensemble size of 20) was selected from the dataset synthesized with
Chapter 2. Clutter Suppression in <strong>Color</strong> <strong>Flow</strong> Imaging: A Hankel-SVD Approach 48<br />
(b) Singular Value Spectrum<br />
Relative Power [dB]<br />
(a) Theoretical Spectrum<br />
80<br />
70<br />
60<br />
50<br />
40<br />
30<br />
20<br />
10<br />
0<br />
-100 0 100 200 300 400 500<br />
Doppler Frequency (f D<br />
) [Hz]<br />
Dopp. Freq. (f D<br />
) [kHz]<br />
Singular Value (σ k<br />
) [dB]<br />
60<br />
50<br />
40<br />
30<br />
20<br />
10<br />
0<br />
1.25<br />
0.75 1<br />
0.5<br />
0.25<br />
-0.25 0<br />
-0.5<br />
-0.75<br />
-1.25<br />
-1<br />
1 2 3 4 5 6 7 8 9 10<br />
Component Order (k)<br />
(c) Mean Frequency Estimates<br />
1 2 3 4 5 6 7 8 9 10<br />
Component Order (k)<br />
Fig. 2-7. Characteristics of principal Hankel components <strong>for</strong> a signal realization in the<br />
arterial flow scenario with 10 cm/s blood velocity: (a) the theoretical Doppler spectrum; (b)<br />
singular value distribution; (c) mean Doppler frequency estimates (marked as ⊗).<br />
arterial flow parameters. The blood velocity used to synthesize this signal realization is 10 cm/s<br />
(i.e. a mean Doppler frequency of 325 Hz). Also, according to the responding parameters listed<br />
in Table 2-3 (bottom of middle column), this signal realization has a peak clutter phase deviation<br />
of 2.6 radians, and it can have time-varying clutter components over the spectral range ±36 Hz.<br />
The theoretical Doppler spectrum <strong>for</strong> this simulation flow scenario is shown in Fig. 2-7a. As can<br />
be seen, the clutter phase modulation gives rise to low-frequency impulses that are closely<br />
spaced in the Doppler spectrum.<br />
Figs. 2-7b and 2-7c respectively show the singular value and mean frequency estimate of<br />
the principal Hankel components that correspond to the selected signal realization. From these<br />
plots, it can be seen that the first principal component corresponds to the clutter eigen-space<br />
based on its distinctly higher singular value magnitude and low mean frequency estimate. Also,<br />
the second principal component appears to belong to the flow eigen-space because its mean<br />
Doppler frequency estimate is close to the blood’s actual mean frequency. The rest of the
Chapter 2. Clutter Suppression in <strong>Color</strong> <strong>Flow</strong> Imaging: A Hankel-SVD Approach 49<br />
(b) Singular Value Spectrum<br />
Relative Power [dB]<br />
80<br />
70<br />
60<br />
50<br />
40<br />
30<br />
20<br />
10<br />
0<br />
(a) Theoretical Spectrum<br />
-200 -100 0 100 200<br />
Doppler Frequency (f D<br />
) [Hz]<br />
Dopp. Freq. (f D<br />
) [Hz]<br />
Singular Value (σ k<br />
) [dB]<br />
70<br />
60<br />
50<br />
40<br />
30<br />
20<br />
10<br />
0<br />
250<br />
200<br />
150<br />
100<br />
50<br />
-50 0<br />
-100<br />
-150<br />
-200<br />
-250<br />
1 2 3 4 5 6 7 8 9 10<br />
Component Order (k)<br />
(c) Mean Frequency Estimates<br />
1 2 3 4 5 6 7 8 9 10<br />
Component Order (k)<br />
Fig. 2-8. Characteristics of principal Hankel components <strong>for</strong> a signal realization in the lowvelocity<br />
flow scenario with 10 cm/s blood velocity. Descriptions are the same as Fig. 2-7.<br />
principal Hankel components appear to correspond to the noise floor since their singular values<br />
are relatively small and their mean frequency estimates do not correspond well with the<br />
predefined spectral characteristics.<br />
Low-Velocity <strong>Flow</strong> Scenario<br />
As a comparison with the arterial flow example, a different realization of Doppler signal<br />
was selected from the low-velocity Doppler dataset. For this realization (also with an ensemble<br />
size of 20), the blood velocity was 10 cm/s, which corresponds to a mean Doppler frequency of<br />
130 Hz. As well, the signal’s peak clutter phase deviation is 10.4 radians, and it has timevarying<br />
clutter components over the spectral range ±114 Hz (see right column of Table 2-3). An<br />
illustration of the Doppler clutter’s wideband spectral characteristics is provided in Fig. 2-8a,<br />
where it can be seen that the low-frequency impulses due to clutter phase modulation are more<br />
significant than the ones in the previous scenario.<br />
Figs. 2-8b and 2-8c show the corresponding singular values and mean frequency<br />
estimates <strong>for</strong> the selected signal’s principal Hankel components. It can be seen that the first two
Chapter 2. Clutter Suppression in <strong>Color</strong> <strong>Flow</strong> Imaging: A Hankel-SVD Approach 50<br />
principal components have distinctly higher singular values and their frequency estimates are<br />
within the clutter spectral range; as such, they appear to correspond to the clutter eigen-space.<br />
On the other hand, the third to fifth principal components appear to belong to the flow subspace<br />
because its mean frequency estimate is near the predefined flow Doppler frequency. As <strong>for</strong> the<br />
sixth to tenth principal components, they appear to be describing the noise floor in view of their<br />
small singular value magnitudes and their spurious mean frequency estimates. <strong>Based</strong> on the<br />
results seen in Figs. 2-7 and 2-8, it can generally be concluded that the clutter and blood<br />
components are likely to be contained within the first few principal Hankel components.<br />
Nevertheless, their actual eigen-space dimensions may vary depending on their respective<br />
Doppler bandwidths and the random variations inherent in each signal realization.<br />
2.5.2 Post-Filter Clutter-to-Blood <strong>Signal</strong> Ratios<br />
Arterial <strong>Flow</strong> Scenario<br />
As a quantitative assessment of the new filter’s per<strong>for</strong>mance in the arterial flow scenario,<br />
the top part of Figs. 2-9 shows the post-filter CBR as a function of the actual blood velocity <strong>for</strong><br />
Doppler ensemble sizes of 10 and 20 samples. The shown results are the 25 th , 50 th , and 75 th -<br />
percentile post-filter CBR estimates obtained from the 10000 signal realizations synthesized at<br />
each blood velocity. The primary observation to be noted from these plots is that the 50 th -<br />
percentile (i.e. median) post-filter CBR curves approached the reciprocal of the expected BSNR<br />
of –10 dB in the higher flow velocity range. This result is expected because, if clutter is<br />
adequately suppressed without distorting the blood signals, the filtered clutter-only signal and the<br />
filtered Doppler signal should respectively be dominated by white noise and blood echoes. As a<br />
comparison <strong>for</strong> these findings, the middle part of Fig. 2-9 shows a similar set of post-filter CBR<br />
curves obtained using a clutter-downmixing filter with similar stopband size; also, the bottom<br />
part of Fig. 2-9 gives a comparison of the median post-filter CBR curves <strong>for</strong> the two filters. It<br />
can be seen that <strong>for</strong> all the shown post-filter CBR curves (i.e. the 25 th , 50 th , and 75 th -percentile<br />
curves), the Hankel-SVD filter has a narrower transition region than the clutter-downmixing<br />
filter in the post-filter CBR curves <strong>for</strong> both Doppler ensemble sizes used in this study. Since a<br />
similar cutoff bandwidth was used <strong>for</strong> the two filters, this result suggests that the Hankel-SVD<br />
filter has a better clutter suppression per<strong>for</strong>mance in the case where the Doppler clutter is<br />
relatively narrowband with respect to the Doppler sampling rate.
Chapter 2. Clutter Suppression in <strong>Color</strong> <strong>Flow</strong> Imaging: A Hankel-SVD Approach 51<br />
(a) N D = 10 (b) N D = 20<br />
Hankel-SVD<br />
Percentile CBR [dB]<br />
0<br />
-5<br />
-10<br />
-15<br />
25th<br />
50th<br />
75th<br />
Percentile CBR [dB]<br />
0<br />
-5<br />
-10<br />
-15<br />
25th<br />
50th<br />
75th<br />
0 4 8 12 16 20 24 28 32 36<br />
Actual Blood Velocity [cm/s]<br />
0 4 8 12 16 20 24 28 32 36<br />
Actual Blood Velocity [cm/s]<br />
Clutter-Downmixing<br />
Percentile CBR [dB]<br />
0<br />
-5<br />
-10<br />
-15<br />
25th<br />
50th<br />
75th<br />
0 4 8 12 16 20 24 28 32 36<br />
Actual Blood Velocity [cm/s]<br />
Percentile CBR [dB]<br />
0<br />
-5<br />
-10<br />
-15<br />
25th<br />
50th<br />
75th<br />
0 4 8 12 16 20 24 28 32 36<br />
Actual Blood Velocity [cm/s]<br />
Filter Comparison<br />
50 th Pecentile CBR [dB]<br />
0<br />
-5<br />
-10<br />
-15<br />
Hankel-SVD<br />
Downmix<br />
0 4 8 12 16 20 24 28 32 36<br />
Actual Blood Velocity [cm/s]<br />
50 th Pecentile CBR [dB]<br />
0<br />
-5<br />
-10<br />
-15<br />
Hankel-SVD<br />
Downmix<br />
0 4 8 12 16 20 24 28 32 36<br />
Actual Blood Velocity [cm/s]<br />
Fig. 2-9. Post-filter CBR as a function of blood velocity in the arterial flow scenario <strong>for</strong><br />
Doppler ensemble sizes of (a) 10 and (b) 20 samples. Shown in the top and middle plots<br />
respectively are the 25 th , 50 th , and 75 th -percentile post-filter CBR estimates <strong>for</strong> the Hankel-<br />
SVD filter and the clutter-downmixing filter. Also, a comparison of the 50 th -percentile postfilter<br />
CBR <strong>for</strong> the two filters is given in the bottom plots. The filter parameters used to obtain<br />
these results are discussed in Section 2.4.2.<br />
Low-Velocity <strong>Flow</strong> Scenario<br />
In contrast to the results <strong>for</strong> the arterial flow scenario, Fig. 2-10 shows the post-filter<br />
CBR curves in the low-velocity flow scenario as a function of the true blood velocity. As can be<br />
expected, the transition regions in all the post-filter CBR curves of both filters are wider in this<br />
case because the Doppler clutter is significantly more wideband in nature (as notable from Fig.
Chapter 2. Clutter Suppression in <strong>Color</strong> <strong>Flow</strong> Imaging: A Hankel-SVD Approach 52<br />
(a) N D = 10 (b) N D = 20<br />
Hankel-SVD<br />
Percentile CBR [dB]<br />
0<br />
-5<br />
-10<br />
-15<br />
25th<br />
50th<br />
75th<br />
Percentile CBR [dB]<br />
0<br />
-5<br />
-10<br />
-15<br />
25th<br />
50th<br />
75th<br />
0 2 4 6 8 10 12 14 16 18<br />
Actual Blood Velocity [cm/s]<br />
0 2 4 6 8 10 12 14 16 18<br />
Actual Blood Velocity [cm/s]<br />
Clutter-Downmixing<br />
Percentile CBR [dB]<br />
0<br />
-5<br />
-10<br />
-15<br />
25th<br />
50th<br />
75th<br />
0 2 4 6 8 10 12 14 16 18<br />
Actual Blood Velocity [cm/s]<br />
Percentile CBR [dB]<br />
0<br />
-5<br />
-10<br />
-15<br />
25th<br />
50th<br />
75th<br />
0 2 4 6 8 10 12 14 16 18<br />
Actual Blood Velocity [cm/s]<br />
Filter Comparison<br />
50 th Pecentile CBR [dB]<br />
0<br />
-5<br />
-10<br />
-15<br />
Hankel-SVD<br />
Downmix<br />
0 2 4 6 8 10 12 14 16 18<br />
Actual Blood Velocity [cm/s]<br />
50 th Pecentile CBR [dB]<br />
0<br />
-5<br />
-10<br />
-15<br />
Hankel-SVD<br />
Downmix<br />
0 2 4 6 8 10 12 14 16 18<br />
Actual Blood Velocity [cm/s]<br />
Fig. 2-10. Post-filter CBR as a function of actual blood velocity in the low-velocity flow<br />
scenario. Descriptions are the same as Fig. 2-9.<br />
2-8a). Nevertheless, within the transition region, the Hankel-SVD filter appears to be able to<br />
attain the steady-state CBR value at lower velocities (near 8-10 cm/s) as compared to the clutterdownmixing<br />
filter (which reaches steady-state CBR near 12-14 cm/s). This result is consistent<br />
with the findings in the arterial flow scenario. Hence, it seems that the Hankel-SVD filter is<br />
generally more capable of suppressing slow-time clutter without concomitantly attenuating the<br />
blood signal components. Another observation to be noted from the plots in Fig. 2-10 is that <strong>for</strong><br />
the case with Doppler ensemble size of 20 samples, the post-filter CBR estimates obtained from<br />
the Hankel-SVD filter appear to have reached a lower steady-state value than the ones found
Chapter 2. Clutter Suppression in <strong>Color</strong> <strong>Flow</strong> Imaging: A Hankel-SVD Approach 53<br />
using the clutter-downmixing filter. This result suggests that, at larger ensemble sizes, the<br />
Hankel-SVD filter may be able to provide a more thorough suppression of the Doppler clutter<br />
than that achievable using a clutter-downmixing filter with the same stopband size. Perhaps a<br />
physical way of perceiving such notion is to recognize that when a larger ensemble size is used,<br />
the Hankel-SVD filter can generally span wideband Doppler clutter into more orthogonal<br />
components because of the increased data resolution, and thus it may be able to remove clutter<br />
more adequately.<br />
2.6 Concluding Remarks<br />
Existing eigen-based clutter filter designs generally require multiple snapshots of Doppler<br />
data that are statistically stationary in order to carry out the eigen-decomposition analysis. To<br />
avoid such requirement, this chapter has presented a new eigen-based filtering strategy – the<br />
Hankel-SVD filter – that is intended to work with the Doppler signal of individual sample<br />
volumes. This new filtering approach can generally be beneficial in vascular imaging studies<br />
where it is difficult to segment out image regions with Doppler sample volumes that have similar<br />
signal characteristics (such as the case where spatially-varying tissue motion is present).<br />
Nevertheless, it may not be as effective as existing eigen-based filters at very small Doppler<br />
ensemble sizes (e.g. when N D = 4) because the Hankel-SVD analysis only decomposes the<br />
Doppler signal into, at most, N D /2 orthogonal basis functions.<br />
A possible limitation of the Hankel-SVD filter is that, during the clutter eigen-space<br />
analysis, Doppler clutter is presumed to be contained in the more dominant Hankel components.<br />
This assumption is generally valid when clutter is significantly higher in strength as compared to<br />
the blood echoes (as is the case in general vascular imaging). However, the assumption would<br />
be invalid if the Doppler clutter and the blood echoes are similar in strength (e.g. in highfrequency<br />
imaging studies where the blood scattering strength is higher). In this case, a more<br />
advanced algorithm is needed to precisely determine the clutter eigen-space dimension. One<br />
possible way of developing such an algorithm is to make use of fuzzy logic and pattern<br />
recognition principles, like the one used by Hu and Gosine (1997) to per<strong>for</strong>m time-series<br />
analysis. Alternatively, it may be possible to use a multi-modal spectral estimator to first extract<br />
a number of principal frequency estimates in parallel and then determine the clutter eigen-space<br />
dimension based on the spectral spread of the estimates.
CHAPTER 3<br />
Velocity Estimation in <strong>Color</strong> <strong>Flow</strong> Imaging:<br />
A Matrix Pencil Approach<br />
3.1 Chapter Overview<br />
It has been mentioned in the introductory chapter that there are two signal processing<br />
challenges when computing velocity estimates from color flow imaging data. The first is related<br />
to the carrier frequency variations that may be present in the received echoes due to physical<br />
mechanisms such as frequency-dependent attenuation and scattering. Given that the carrier<br />
frequency is one of the factors in the Doppler equation, these variations may lead to biases in the<br />
Doppler frequency estimates. The second challenge is associated with the non-idealities in the<br />
transfer response of the clutter suppression filter. As flow estimates are often derived from the<br />
filtered Doppler signal, these filter non-idealities may also give rise to Doppler frequency biases<br />
during flow estimation. Between the two signal processing challenges, the one due to carrier<br />
distortions is less significant in practice because the transmit pulses used in color flow imaging<br />
are relatively narrowband in nature (see Section 1.4.2) and are inherently less prone to carrier<br />
frequency shifts. On the other hand, biases due to clutter filter non-idealities can be significant<br />
depending on the filter characteristics and the Doppler spectral contents of blood echoes.<br />
To address the biasing problems originating from the clutter filtering operation, this<br />
chapter presents a new parametric velocity estimation framework called the Matrix Pencil. This<br />
new estimator works by exploiting the properties of an algebraic <strong>for</strong>m known as matrix pencil<br />
and in turn treating flow estimation as a generalized eigenvalue (GE) problem. To <strong>for</strong>mulate<br />
discussion on the Matrix Pencil, the rest of this chapter is organized as follows:<br />
• Section 3.2 reviews the principles of existing flow estimation strategies and discusses their<br />
advantages and limitations;<br />
• Section 3.3 presents the theoretical <strong>for</strong>mulation of the Matrix Pencil estimation framework<br />
and describes its applications in color flow data processing;<br />
- 54 -
Chapter 3. Velocity Estimation in <strong>Color</strong> <strong>Flow</strong> Imaging: A Matrix Pencil Approach 55<br />
• Sections 3.4 and 3.5 discuss the simulation approach used to analyze the new estimator’s<br />
per<strong>for</strong>mance and show the corresponding results obtained from this study;<br />
• Section 3.6 summarizes the highlights of the new flow estimation framework and its<br />
applicability to color flow imaging.<br />
3.2 Existing <strong>Flow</strong> Estimation Strategies<br />
3.2.1 Non-Parametric Estimators<br />
Fundamental Considerations<br />
In color flow imaging, a two-stage process is often used to compute velocity estimates.<br />
First, a clutter filter is applied to the Doppler signal of each sample volume to remove lowfrequency<br />
echoes originating from tissues and vessel walls. The filtered signal is then passed<br />
into a non-parametric flow estimator to compute the mean flow velocity. For this two-stage<br />
estimation approach, it is presumed that the clutter filter can adequately suppress clutter so that<br />
the filtered Doppler data is only consisted of blood echoes and filtered white noise. <strong>Based</strong> on<br />
such notion, the filtered Doppler data vector y (of length N D ) can be described by the following<br />
signal model:<br />
T<br />
[ ( 0), y(1),<br />
, y(<br />
D<br />
−1)<br />
] = b w<br />
f<br />
y = y K N + , (3-1)<br />
where y(n) is the n th filtered sample, while b and w f respectively denote the signal vectors <strong>for</strong><br />
blood echoes and filtered white noise. This signal model often serves as the starting point in the<br />
derivation of various non-parametric mean velocity estimators.<br />
Frequency-<strong>Based</strong> <strong>Methods</strong><br />
One approach <strong>for</strong> finding the mean flow velocity is to first estimate the mean frequency<br />
of the filtered Doppler signal and then use the Doppler equation to convert the frequency<br />
estimate into a velocity value. In this estimation approach, computation of the mean Doppler<br />
frequency can theoretically be per<strong>for</strong>med by dividing between the first-order and zeroth-order<br />
Doppler spectral moments. Alternatively, by making use of the Wiener-Khinchin relation † , the<br />
same computations can be carried out in a more efficient manner through an analysis of the<br />
† The Wiener-Khinchin relation states that the autocorrelation function of a signal is equivalent to the inverse<br />
Fourier trans<strong>for</strong>m of the signal’s power spectral density.
Chapter 3. Velocity Estimation in <strong>Color</strong> <strong>Flow</strong> Imaging: A Matrix Pencil Approach 56<br />
Doppler autocorrelation function (i.e. without directly analyzing the Doppler spectrum). As first<br />
described by Kasai et al. (1985) in the context of color flow imaging, the autocorrelation-based<br />
approach would lead to the following expression <strong>for</strong> the mean Doppler frequency:<br />
f<br />
D(est)<br />
1<br />
=<br />
2πT<br />
PRI<br />
arg<br />
ND<br />
{ } ∑ −<br />
R (1) <strong>for</strong> R (1) =<br />
y<br />
y<br />
1<br />
n=<br />
1<br />
y(<br />
n)<br />
y<br />
∗<br />
( n −1)<br />
, (3-2)<br />
where R y (1) is the Doppler autocorrelation function evaluated at a lag of one sample. <strong>Based</strong> on<br />
this equation, it can be seen that the lag-one autocorrelation phase is the primary factor in the<br />
mean Doppler frequency estimate. The phase quantity, however, can be affected by signal<br />
aliasing due to the sampled nature of the Doppler data (see Section 1.4.2), and such problem is<br />
often considered as a theoretical limitation of the autocorrelation estimator. Nonetheless, the<br />
aliasing limit is actually appreciated in practice because the unaliased velocity range is often<br />
used to define the mapping scale when designing the color map. Also, as pointed out by Tamura<br />
et al. (1991), aliasing is sometimes introduced intentionally in color flow images to improve the<br />
visualization of laminar flow patterns.<br />
The estimation variance of the lag-one autocorrelator shown in (3-2) can be reduced by<br />
expanding the Doppler autocorrelation into a two-dimensional (2D) function based on multiple<br />
Doppler signal snapshots within a given depth range (Torp et al. 1994). However, its use in<br />
color flow data processing is often not vital because the estimation variance can also be reduced<br />
via spatial averaging of the velocity estimates from adjacent sample volumes within the imaging<br />
view. As pointed out by Loupas et al. (1995a, 1995b), another possible advantage of the 2D<br />
autocorrelation approach is that it can be used to jointly estimate both the mean Doppler<br />
frequency and the instantaneous carrier frequency. Such feature is useful <strong>for</strong> correcting Doppler<br />
frequency biases originating from carrier frequency variations that arise when wideband pulses<br />
are used to acquire color flow data (even though wideband data acquisition is not common in<br />
color flow imaging as we noted in Section 1.4.2).<br />
Time-Shift-<strong>Based</strong> <strong>Methods</strong><br />
Another approach <strong>for</strong> estimating the mean flow velocity is to find the average inter-pulse<br />
time shift between post-filter echo envelopes of the same beam line and then apply (1-2) to<br />
convert the time shift estimate into a velocity value. As reviewed by Alam and Parker (2003),<br />
this time-shift-based estimation approach is essentially derived from target tracking principles
Chapter 3. Velocity Estimation in <strong>Color</strong> <strong>Flow</strong> Imaging: A Matrix Pencil Approach 57<br />
that are used in radar and sonar. In terms of its implementation, time shift estimation can be<br />
per<strong>for</strong>med in two different ways. The first way, as initially applied to color flow imaging by<br />
Bonnefous and Pesque (1986), involves finding the average time lag that yields the maximum<br />
cross-correlation between successive echo envelopes of the beam line. The second way, as<br />
proposed by Ferrara and Algazi (1991a, 1991b) as well as Alam and Parker (1995), involves<br />
computing a time-shift likelihood function from geometric projection principles and then<br />
searching <strong>for</strong> its location of maximum. For both methods of implementation, their per<strong>for</strong>mance<br />
is not affected by carrier frequency distortions because inter-pulse time shifts are not physically<br />
dependent on the carrier frequency (as seen in (1-2)). They also do not suffer from aliasing<br />
problems since they are not based on the analysis of inter-pulse phase changes. However, the<br />
theoretical advantages offered by these estimators appear to be non-essential to color flow data<br />
processing (e.g. aliasing is indeed preferred in color flow imaging as noted earlier). As such,<br />
time-shift-based estimators are often not used in color flow imaging, even though they have been<br />
a significant part of early research ef<strong>for</strong>ts on color flow signal processing.<br />
3.2.2 Parametric Estimators<br />
Fundamental Considerations<br />
In using non-parametric estimators to find flow velocities, estimation biases can be<br />
expected whenever the clutter filter distorts parts of the blood signal or fails to adequately<br />
suppress clutter. An illustration of this problem <strong>for</strong> a frequency-based non-parametric estimator<br />
is given in Figure 3-1a. Note that, at low blood-signal-to-noise ratios (BSNR), further estimation<br />
biases can be anticipated because the filtered white noise (i.e. colored noise) becomes a<br />
significant part of the filtered Doppler signal and in turn adds bias to the flow estimates. One of<br />
the ways to account <strong>for</strong> these clutter filter biases is to apply a post facto correction factor to the<br />
flow estimates. For instance, Rajaonah et al. (1994) have reported a frequency-based bias<br />
correction scheme that applies a correction factor based on the mean Doppler frequency of<br />
colored noise samples. Their approach, however, is intended to work with time-invariant clutter<br />
filters such as FIR filters, and thus it remains unclear as to whether this method can sufficiently<br />
account <strong>for</strong> biases due to time-variant filters like regression filters and initialized IIR filters.<br />
In order to effectively account <strong>for</strong> clutter filter biases, it is theoretically advantageous to<br />
use estimation strategies that can be directly applied to the raw (i.e. unfiltered) Doppler signal.
Chapter 3. Velocity Estimation in <strong>Color</strong> <strong>Flow</strong> Imaging: A Matrix Pencil Approach 58<br />
(a) Non-Parametric Spectral Estimator<br />
Clutter<br />
Spectrum<br />
Filter<br />
Stopband<br />
0<br />
Mean <strong>Flow</strong><br />
Freq (Biased)<br />
Blood<br />
Spectrum<br />
f D<br />
(b) Parametric Spectral Estimator<br />
Clutter<br />
Spectrum<br />
Clutter<br />
Mode<br />
0<br />
<strong>Flow</strong> <strong>Signal</strong><br />
Mode<br />
Blood<br />
Spectrum<br />
f D<br />
Fig. 3-1. Difference between the spectral estimates of (a) non-parametric and (b) rank-two<br />
parametric flow estimators. Note that, in this example, the mean estimate in (a) is biased<br />
away from the blood spectral peak because of filter distortions, while the two spectral modes<br />
in (b) are relatively unbiased.<br />
This rationale has motivated the development of parametric estimators that work by analyzing<br />
the principal Doppler spectral contents (i.e. they are frequency-based estimators). For these<br />
estimators, their principles are generally related to the following principal-component signal<br />
model:<br />
T<br />
[ x( 0), x(1),<br />
, x(<br />
N −1<br />
] = b + c + w ≈ χ v w<br />
x = K<br />
D<br />
) ∑ k CS( k )<br />
+ . (3-3a)<br />
In the above expression, x(n) is the n th raw Doppler data sample, while x, b, c, and w respectively<br />
denote the length-N D vectors <strong>for</strong> raw Doppler signal, blood echoes, clutter, and white noise; also,<br />
K is the number of principal components in the signal model, whereas χ k and v CS(k) are the<br />
weight and complex sinusoid vector of the k th component. Note that v CS(k) essentially has the<br />
following vector <strong>for</strong>m:<br />
CS(<br />
k )<br />
K<br />
k = 1<br />
2 ( ND −1)<br />
2 D, PRI<br />
[ 1, , , , ] T<br />
j πf<br />
kT<br />
z z K z <strong>for</strong> z = e<br />
v =<br />
, (3-3b)<br />
k<br />
k<br />
k<br />
where f D,k is the k th principal Doppler frequency. It should be pointed out that, from a subspace<br />
perspective, this principal-component signal model is equivalent to the eigen-structure of a raw<br />
Doppler signal whose rank is equal to K. An illustration of the principal frequencies found using<br />
a rank-two parametric estimator is shown in Figure 3-1b.<br />
Autoregressive Modeling<br />
Autoregressive (AR) modeling (also known as linear prediction) is a type of parametric<br />
estimation approach that computes the principal Doppler frequencies through an all-pole signal<br />
k
Chapter 3. Velocity Estimation in <strong>Color</strong> <strong>Flow</strong> Imaging: A Matrix Pencil Approach 59<br />
analysis. As reviewed by Vaitkus and Cobbold (1988) in the context of Doppler ultrasound, this<br />
approach begins by expressing each Doppler data sample as a linear combination of white noise<br />
and past data samples. In particular, <strong>for</strong> a K th -order AR model, the n th sample of the Doppler<br />
signal x(n) is mathematically described by the following difference equation:<br />
K<br />
∑<br />
x( n)<br />
= w(<br />
n)<br />
− c x(<br />
n − k)<br />
, (3-4a)<br />
k = 1<br />
where w(n) and c k are respectively the n th white noise sample and the k th model fitting<br />
coefficient. Note that this difference equation has the following all-pole <strong>for</strong>m in the discrete<br />
frequency domain:<br />
X<br />
1<br />
K<br />
−1<br />
AR<br />
( f<br />
D<br />
) =<br />
= ( z z )<br />
−1<br />
−K<br />
∏ −<br />
k<br />
1+<br />
c1z<br />
+ K + cK<br />
z k = 1<br />
k<br />
j 2πf<br />
DT<br />
z=<br />
e PRI<br />
, (3-4b)<br />
where z k is the k th characteristic mode (i.e. the k th root of the denominator polynomial). From<br />
discrete-time signal theory, it is well-known that the difference equation’s natural response takes<br />
on the same <strong>for</strong>m as the principal-component signal model given in (3-3a) when no repeated<br />
modes are present. As such, the characteristic modes of the difference equation can be used to<br />
estimate the principal frequencies of the Doppler signal. Specifically, since z is defined as<br />
exp{j2πf D T PRI }, the k th principal Doppler frequency can be found from the k th characteristic mode<br />
as follows:<br />
f<br />
1<br />
= arg . (3-5a)<br />
{ }<br />
D, k<br />
z k<br />
2πTPRI<br />
In turn, the modal flow frequency can be identified from the principal spectral estimate that has<br />
the largest magnitude:<br />
{ f ∀ k ∈[1,<br />
]}<br />
f<br />
D(est)<br />
= f<br />
D, k<br />
<strong>for</strong> k<br />
b<br />
= arg max<br />
b D,<br />
k<br />
K . (3-5b)<br />
In practice, the coefficients c k in the AR model are the main parameters that characteristic<br />
modes depend on, and they are typically found by fitting the Doppler samples onto the data<br />
model. As discussed in spectral analysis textbooks (e.g. see Stoica and Moses 1997, Ch. 3), an<br />
effective way of optimizing these coefficients is to minimize the mean-squared fitting error from<br />
both <strong>for</strong>ward and backward regression perspectives. Such a least-squares fit can be computed<br />
using the Prony <strong>for</strong>ward-backward fitting method (also known as modified covariance method),<br />
which is described by the following system of equations given a Doppler ensemble size of N D :<br />
k
Chapter 3. Velocity Estimation in <strong>Color</strong> <strong>Flow</strong> Imaging: A Matrix Pencil Approach 60<br />
⎡ x(<br />
K −1)<br />
⎢<br />
x(<br />
K)<br />
⎢<br />
⎢ M<br />
⎢<br />
⎢ x(<br />
N<br />
D<br />
− 2)<br />
⎢<br />
∗<br />
x (1)<br />
⎢<br />
∗<br />
⎢ x (2)<br />
⎢ M<br />
⎢<br />
∗<br />
⎢⎣<br />
x ( N<br />
D<br />
− K)<br />
x(<br />
K − 2)<br />
x(<br />
K −1)<br />
x(<br />
N<br />
x<br />
x<br />
∗<br />
x ( N<br />
D<br />
∗<br />
D<br />
∗<br />
M<br />
− 3)<br />
(2)<br />
(3)<br />
M<br />
− K + 1)<br />
L<br />
L<br />
O<br />
L<br />
L<br />
L<br />
O<br />
L<br />
x(0)<br />
⎤ ⎡<br />
x(1)<br />
⎥ ⎢<br />
⎥ ⎢<br />
M ⎥⎡<br />
c ⎢<br />
1 ⎤<br />
⎥ ⎥ ⎢<br />
x(<br />
N<br />
⎢<br />
D<br />
− K −1)<br />
⎥⎢<br />
c2<br />
⎥ = −⎢<br />
∗<br />
x ( K)<br />
⎥⎢<br />
M ⎥ ⎢<br />
∗<br />
⎥⎢<br />
⎥ ⎢<br />
x ( K + 1) ⎥⎣cK<br />
⎦ ⎢<br />
M ⎥ ⎢<br />
⎥ ⎢<br />
∗<br />
x ( N<br />
D<br />
−1)<br />
⎥⎦<br />
⎢⎣<br />
x<br />
∗<br />
x(<br />
K)<br />
⎤<br />
x(<br />
K + 1)<br />
⎥<br />
⎥<br />
M ⎥<br />
⎥<br />
x(<br />
N<br />
D<br />
−1)<br />
⎥ . (3-6)<br />
∗<br />
x (0) ⎥<br />
∗<br />
⎥<br />
x (1) ⎥<br />
M ⎥<br />
⎥<br />
( N<br />
D<br />
− K −1)<br />
⎥⎦<br />
In this system of equations, the upper and lower halves respectively account <strong>for</strong> the <strong>for</strong>ward and<br />
backward regressions of the Doppler data. Note that these equations may be recursively solved<br />
via an approach known as the Burg method.<br />
In the context of color flow data processing, Loupas and McDicken (1990) first<br />
considered the use of AR modeling in flow estimation. They showed that a first-order AR<br />
estimator, which only finds one Doppler spectral mode, is essentially equivalent to the lag-one<br />
autocorrelator shown in (3-2). Subsequently, Ahn and Park (1991) attempted to use AR<br />
modeling in flow estimation studies that work with raw Doppler data. In particular, they<br />
developed a second-order AR estimator to simultaneously find the principal Doppler frequency<br />
of clutter and blood echoes. However, in this estimator, clutter is inherently assumed to be a<br />
single low-frequency complex sinusoid. Such an assumption may not always be valid, especially<br />
in cases with wideband clutter.<br />
Multiple <strong>Signal</strong> Classification<br />
Another way of computing principal Doppler frequencies is to per<strong>for</strong>m eigen-analysis on<br />
the correlation statistics of the Doppler signal. In particular, a <strong>for</strong>m of eigen-analysis called<br />
multiple signal classification (MUSIC) has demonstrated potential in obtaining flow estimates<br />
from the raw Doppler data. As first applied to Doppler ultrasound by Allam and Greenleaf<br />
(1996), the MUSIC approach begins by decomposing the Doppler signal into a set of orthogonal<br />
basis functions though an eigen-decomposition of the correlation matrix (just like the eigenregression<br />
filter as described in Section 2.2.2). For a Doppler ensemble size N D , such<br />
decomposition can be described by the following set of expressions † :<br />
† This decomposition, sometimes referred to as the Karhunen-Loeve (KL) expansion, is essentially the same set of<br />
equations given in (2-9) of Section 2.2.2. It is restated here <strong>for</strong> convenience of discussion.
Chapter 3. Velocity Estimation in <strong>Color</strong> <strong>Flow</strong> Imaging: A Matrix Pencil Approach 61<br />
D<br />
∑<br />
H<br />
H<br />
R = E{ xx = λ e e , (3-7a)<br />
x<br />
}<br />
ND<br />
⎧λk<br />
, k = l<br />
x = ∑γ ke<br />
k<br />
<strong>for</strong> E{ γ<br />
kγ<br />
l}<br />
= ⎨ , (3-7b)<br />
k = 1<br />
⎩ 0 , k ≠ l<br />
where R x is the correlation matrix of the Doppler signal, while γ k , λ k , and e k are the respective<br />
expansion coefficient, eigenvalue, and eigenvector of the k th orthogonal basis function. In<br />
practice, the Doppler correlation matrix in (3-7a) can be computed through ensemble averaging<br />
of multiple Doppler snapshots over a range of depth with similar signal characteristics.<br />
Within the eigen-decomposition given in (3-7), the Doppler signal contents are mainly<br />
contained in the high-energy basis functions since their strength is generally greater than<br />
background white noise. As such, the basis functions of the decomposition can be separated into<br />
signal and noise components, thereby <strong>for</strong>ming two mutually orthogonal subspaces. From this<br />
signal separation, a frequency pseudo-spectrum can then be computed by finding the reciprocal<br />
of the cross-correlation between noise components and various complex sinusoids (with spectral<br />
peaks appearing at frequencies where the cross-correlation approaches zero). Specifically, <strong>for</strong> a<br />
Doppler signal with K principal bases, the pseudo-spectrum X MUSIC (f D ) can be expressed as:<br />
X<br />
f<br />
N<br />
k = 1<br />
⎡ | |<br />
1<br />
1<br />
) = =<br />
<strong>for</strong> E =<br />
⎢<br />
n<br />
⎢<br />
e<br />
2 H H<br />
K + 1<br />
e<br />
K +<br />
H<br />
En<br />
v v<br />
CS<br />
CSEnEn<br />
v<br />
CS<br />
⎢⎣<br />
| |<br />
MUSIC<br />
(<br />
D<br />
2<br />
D<br />
k<br />
k<br />
k<br />
| ⎤<br />
L e<br />
⎥<br />
N<br />
⎥<br />
, (3-8a)<br />
| ⎥⎦<br />
where E n is the noise subspace matrix (which consists of the N D –K+1 least dominant<br />
eigenvectors) and v CS is a length-N D complex sinusoid vector of frequency f D . As shown in<br />
spectral analysis textbooks (e.g. Stoica and Moses 1997, Sec. 4.5), the denominator in (3-8a) is<br />
essentially equal to the following (2N D –1) th -order polynomial whose coefficients are given by the<br />
sum of elements along each diagonal of the matrix E n E H n :<br />
E<br />
H<br />
n<br />
v<br />
CS<br />
2<br />
= c<br />
ND<br />
−1<br />
z<br />
ND<br />
−1<br />
+ c<br />
ND<br />
−2<br />
z<br />
ND<br />
−2<br />
+ K + c<br />
−(<br />
ND<br />
−1)<br />
z<br />
−(<br />
ND<br />
−1)<br />
<strong>for</strong> c<br />
k<br />
j 2πf<br />
z=<br />
e DTPRI<br />
=<br />
N − 1N<br />
−1<br />
D D<br />
∑∑<br />
i=<br />
0<br />
j=<br />
0<br />
δ<br />
k = j−i<br />
. (3-8b)<br />
e ( i,<br />
j)<br />
n<br />
In the above expression, c k is the k th polynomial coefficient, e n (i, j) is the entry of E n E n H at row i<br />
and column j, and δ k=j–i is a Kronecker delta function that equals to one if k = j–i and equals to<br />
zero otherwise. Note that the polynomial only has K non-spurious modes.
Chapter 3. Velocity Estimation in <strong>Color</strong> <strong>Flow</strong> Imaging: A Matrix Pencil Approach 62<br />
Table 3-1. Comparison of Existing Frequency-<strong>Based</strong> Velocity Estimators<br />
Type Advantage Limitation<br />
Lag-one<br />
Autocorrelator<br />
AR Estimator<br />
MUSIC<br />
Simple to implement with low<br />
computation demand<br />
Finds flow velocity directly from<br />
raw Doppler data without clutter<br />
filtering<br />
Same as AR estimator, but more<br />
resilient to white noise<br />
Only works on filtered Doppler data;<br />
prone to biases from filter distortions<br />
Least-squares fit of c k fails at high<br />
white noise levels; efficacy depends<br />
on choice of model order<br />
Needs multiple data snapshots to<br />
find correlation matrix; efficacy<br />
depends on eigen-structure rank<br />
Like AR modeling, the eigen-modes in the MUSIC pseudo-spectrum can be found<br />
numerically by solving <strong>for</strong> the denominator roots, and correspondingly the modal flow frequency<br />
can be found from the principal frequency estimates using (3-5). <strong>Based</strong> on this root-finding<br />
principle, Vaitkus and Cobbold (1998) developed a closed-<strong>for</strong>m parametric estimator called<br />
Root-MUSIC to estimate principal flow velocities from raw Doppler data. Their approach,<br />
which uses a rank-two eigenstructure to model the raw Doppler signal, was analyzed using in<br />
vivo Doppler data whose clutter can be sufficiently modeled as a single complex sinusoid<br />
(Vaitkus et al. 1998). Note that another way to solve <strong>for</strong> the MUSIC eigen-modes is to use a<br />
peak searching algorithm to find the principal peaks in the pseudo-spectrum. Such approach was<br />
used by Allam et al. (1996) to process Doppler data acquired from a string phantom and a<br />
reflective surface that respectively simulate blood flow and stationary clutter.<br />
3.3.2 Comparison of Frequency-<strong>Based</strong> Velocity Estimators<br />
Table 3-1 summarizes the advantages and limitations of the three frequency-based<br />
velocity estimators covered in this section. As indicated, the lag-one autocorrelator has the<br />
advantage of being computationally efficient. Indeed, the low computing burden of this<br />
estimator is well appreciated in early research and developments because of processing power<br />
limitations. However, as already pointed out, the lag-one autocorrelator can only be applied to<br />
filtered Doppler data, and thus, its velocity estimates are inherently prone to biases from clutter<br />
filter distortions.<br />
Contrary to the lag-one autocorrelator, the AR estimator has the theoretical advantage of<br />
being able to obtain velocity estimates by directly processing the raw Doppler data. This
Chapter 3. Velocity Estimation in <strong>Color</strong> <strong>Flow</strong> Imaging: A Matrix Pencil Approach 63<br />
estimator first uses a least-squares fitting procedure to compute the characteristic modes of the<br />
raw Doppler signal and then identifies the modal flow component based on the largest-frequency<br />
mode, thereby avoiding the need <strong>for</strong> clutter filtering. Nevertheless, the primary shortcoming of<br />
the AR estimator is that the least-squares fitting solution to the model coefficients assumes that<br />
the Doppler data samples are free of noise perturbations. Thus, its accuracy tends to degrade<br />
significantly as the noise level increases. Another limitation with the AR estimator is that its<br />
flow estimation per<strong>for</strong>mance is dependent on the choice of the model order (or eigen-structure<br />
rank). Specifically, a wrong choice of model order would give rise to spurious spectral modes<br />
and in turn give incorrect modal velocity estimates.<br />
Unlike the AR estimator, the MUSIC estimator is more resilient to white noise because it<br />
first makes use of an eigen-decomposition procedure to separate the signal bases from noise floor<br />
components. There<strong>for</strong>e, when applied to raw Doppler data with high noise levels, this estimator<br />
can obtain modal velocity estimates that are less biased than the ones found from AR modeling.<br />
On a different note, it is worth pointing out that the <strong>for</strong>mulation of MUSIC begins with the same<br />
eigen-decomposition step as seen <strong>for</strong> the eigen-regression filter. <strong>Based</strong> on this notion, it follows<br />
that MUSIC has limitations similar to the eigen-regression filter. In particular, MUSIC would<br />
need multiple signal snapshots that are statistically stationary in order to <strong>for</strong>m an accurate<br />
estimate of the Doppler correlation matrix. One further limitation of the MUSIC approach is<br />
that, as similar to AR modeling, the per<strong>for</strong>mance of this estimator is contingent upon suitable<br />
choice of the eigen-structure rank. As such, an adaptive rank selection algorithm is needed in<br />
order <strong>for</strong> this estimator to be effective in general.<br />
3.3 The Matrix Pencil Estimator<br />
3.3.1 Background Considerations<br />
Design Motivations<br />
From the previous work on AR modeling and MUSIC, it can be seen that parametric<br />
estimation strategies have potential in obtaining modal flow estimates in the presence of clutter.<br />
However, as noted earlier, these parametric estimators have certain individual limitations that<br />
reduce their efficacy in color flow data processing. For instance, the least-squares fitting<br />
procedure used by an AR estimator to compute the model coefficients is only effective at high
Chapter 3. Velocity Estimation in <strong>Color</strong> <strong>Flow</strong> Imaging: A Matrix Pencil Approach 64<br />
blood-signal-to-noise ratios. As <strong>for</strong> MUSIC, its multi-snapshot way of estimating the Doppler<br />
correlation matrix inherently requires statistical stationarity amongst the Doppler signal vectors<br />
over the specified range of depth, and such requirement cannot always be satisfied due to the<br />
spatially varying nature of clutter and flow dynamics. In view of these theoretical limitations,<br />
we considered the development of a new parametric flow estimator that is resilient to white noise<br />
and that does not require multiple Doppler signal ensembles.<br />
Fundamental Principles<br />
The new parametric estimation approach to be presented in this section is based on a<br />
framework originally developed <strong>for</strong> pole estimation in control systems as well as direction-ofarrival<br />
estimation in array processing (Hua and Sarkar 1990). It works by exploiting the<br />
properties of an algebraic <strong>for</strong>m called matrix pencil, and hence it will be referred to as the Matrix<br />
Pencil estimator. In terms of its principles, this new estimator is primarily concerned with the<br />
solution to the following generalized eigenvalue (GE) problem:<br />
A q = λ A q ⇔ A − A ) q = 0 . (3-9)<br />
1 0<br />
(<br />
1<br />
λ<br />
0<br />
In this equation, A 1 and A 0 are singular matrices of the same dimensions, and q is the<br />
generalized eigenvector <strong>for</strong> a particular eigenvalue λ. As first extensively discussed by<br />
Gantmacher (1960, Ch. XII), the set of matrices A 1 –λA 0 created from all values of λ is referred<br />
to as a matrix pencil † . The GEs of this matrix pencil are the particular values of λ that yield nonzero<br />
solutions to the eigenvector q in (3-9). Such values can also be regarded as those that<br />
reduce the rank of the matrix pencil so that a non-empty nullspace exists <strong>for</strong> A 1 –λA 0 .<br />
3.3.2 Theoretical Formulation<br />
Construction of Data Matrix Pair<br />
The overall <strong>for</strong>mulation of the Matrix Pencil estimator can be summarized in the<br />
flowchart given in Fig. 3-2. In terms of its details, this estimator begins by constructing the<br />
matrix pair in (3-9) in a way so that its GEs contain in<strong>for</strong>mation about the principal Doppler<br />
frequencies. One particular way to <strong>for</strong>m the matrix pair is to define A 1 and A 0 such that their<br />
structure only differs from each other by a diagonal matrix operator (Hua and Sarkar 1990).<br />
† In mathematics, the term pencil generally refers to a set of entities that share a common property, such as passage<br />
through the same given point.
Chapter 3. Velocity Estimation in <strong>Color</strong> <strong>Flow</strong> Imaging: A Matrix Pencil Approach 65<br />
Doppler<br />
<strong>Signal</strong><br />
Create Hankel<br />
Matrix<br />
A =<br />
0 1<br />
2<br />
1<br />
2<br />
3<br />
2<br />
3<br />
3<br />
4<br />
4<br />
5<br />
4 5 6<br />
5 6 7<br />
Per<strong>for</strong>m SVD<br />
P<br />
A = Σσ k<br />
u k<br />
v k<br />
H<br />
1<br />
Obtain Filtered Data<br />
Matrix<br />
Ã=<br />
~ ~ ~ 0 1 2 ~ ~ ~ 1 2 3 ~ ~ ~ 2 3 4 ~ ~ ~ 3 4 5 ~ ~ ~ 4 5 6<br />
~ ~ ~ 5 6 7<br />
<strong>Flow</strong><br />
Frequency<br />
Estimate<br />
f D(est)<br />
Find Largest<br />
Frequency<br />
f D(est) = max{|f D,k |}<br />
Find Generalized<br />
<strong>Eigen</strong>values<br />
P<br />
à à = Σλ e e H<br />
0+ 1 k k k<br />
1<br />
Create Matrix Pencil<br />
~ ~ ~ 1 2 3 ~ ~ ~ 2 3 4 ~ ~ ~ 3 4 5 ~ ~ ~ 4 5 6<br />
à 1 =<br />
à 0 =<br />
~ ~ ~ 5 6 7<br />
~ 5<br />
~ 6<br />
~ 7<br />
~ ~ ~ 0 1 2 ~ ~ ~ 1 2 3 ~ ~ ~ 2 3 4 ~ ~ ~ 3 4 5 ~ ~ ~ 4 5 6<br />
Fig. 3-2. Block diagram overview of the Matrix Pencil estimator with low-rank filtering.<br />
During operation, this estimator is applied to the Doppler signal of each sample volume.<br />
Specifically, <strong>for</strong> a Doppler ensemble size of N D (indexed from 0 to N D –1), A 1 and A 0 can be<br />
defined as the following Hankel data matrices:<br />
⎡ x(1)<br />
x(2)<br />
L x(<br />
P)<br />
⎤<br />
⎢<br />
x<br />
x<br />
x P<br />
⎥<br />
⎢<br />
(2) (3) L ( + 1)<br />
A ⎥<br />
1<br />
=<br />
, (3-10a)<br />
⎢ M<br />
M O M ⎥<br />
⎢<br />
⎥<br />
⎣x(<br />
N<br />
D<br />
− P)<br />
x(<br />
N<br />
D<br />
− P + 1) L x(<br />
N<br />
D<br />
−1)<br />
⎦<br />
( N −P)<br />
× P<br />
⎡ x(0)<br />
x(1)<br />
L x(<br />
P −1)<br />
⎤<br />
⎢<br />
x<br />
x<br />
x P<br />
⎥<br />
⎢<br />
(1) (2) L ( )<br />
A ⎥<br />
0<br />
=<br />
. (3-10b)<br />
⎢ M<br />
M O M ⎥<br />
⎢<br />
⎥<br />
⎣x(<br />
N<br />
D<br />
− P −1)<br />
x(<br />
N<br />
D<br />
− P)<br />
L x(<br />
N<br />
D<br />
− 2) ⎦<br />
D<br />
( N −P)<br />
× P<br />
An example on the construction of this matrix pair is given in Fig. 3-3. Note that, in A 1 and A 0 ,<br />
P is a pencil dimension parameter that is similar to the one defined <strong>for</strong> the full data matrix given<br />
in (2-12) of Chapter 2; <strong>for</strong> a K th -order signal eigen-structure, this parameter must be greater than<br />
K but less than ⎡N D /2⎤ (where the ⎡.⎤ operator denotes the smallest integer greater than or equal<br />
to N D /2). It should be also pointed out that the two matrices defined in (3-10) are essentially<br />
overlapping subsets of the full data matrix A defined in (2-12). Hence, like the full data matrix,<br />
they assume that the samples within the Doppler signal vector are statistically stationary.<br />
To appreciate how the matrix pencil A 1 –λA 0 gives GEs that correspond to the principal<br />
frequencies, it is necessary to consider the decomposition properties of the two data matrices. In<br />
D
Chapter 3. Velocity Estimation in <strong>Color</strong> <strong>Flow</strong> Imaging: A Matrix Pencil Approach 66<br />
1 2 3<br />
2 3 4<br />
3<br />
4<br />
4<br />
5<br />
5<br />
6<br />
5 6 7<br />
Data matrix A 1<br />
Doppler signal vector x<br />
0<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
Data matrix A 0<br />
0 1 2<br />
1 2 3<br />
2<br />
3<br />
3<br />
4<br />
4<br />
5<br />
4 5 6<br />
Fig. 3-3. Illustration of how to <strong>for</strong>m the data matrices A 1 and A 0 in the Matrix Pencil<br />
estimator. In this example, the case where N D =8 and P=3 is assumed.<br />
particular, with reference to the signal model given in (3-3), the data matrices A 1 and A 0 in the<br />
absence of noise can be decomposed into the following <strong>for</strong>ms:<br />
A = , (3-11a)<br />
1<br />
Z<br />
LDΦ<br />
Z<br />
R<br />
A = , (3-11b)<br />
0<br />
Z<br />
LDZ<br />
R<br />
where the matrices Z L , Z R , D, and Φ are defined as follows:<br />
⎡<br />
⎢<br />
⎢<br />
⎢<br />
⎢<br />
⎣z<br />
1<br />
2<br />
K<br />
Z<br />
L<br />
=<br />
, (3-12a)<br />
( N<br />
1<br />
1 1 L<br />
z<br />
D<br />
M<br />
−P−1)<br />
z<br />
( N<br />
2<br />
z<br />
D<br />
M<br />
−P−1)<br />
L<br />
O<br />
L<br />
z<br />
( N<br />
K<br />
1 ⎤<br />
z<br />
D<br />
M<br />
−P−1)<br />
⎥ ⎥⎥⎥ ⎦<br />
( N<br />
−P)<br />
× K<br />
z k = e<br />
D<br />
j 2πf<br />
D,kTPRI<br />
⎡1<br />
( P−1)<br />
1<br />
1<br />
⎢<br />
( P−1)<br />
⎥<br />
⎢1<br />
z2<br />
L z2<br />
Z ⎥<br />
R<br />
=<br />
, (3-12b)<br />
⎢M<br />
⎢<br />
⎢⎣<br />
1<br />
z<br />
z<br />
M<br />
K<br />
L<br />
O<br />
L<br />
z<br />
z<br />
M<br />
( P−1)<br />
K<br />
⎤<br />
⎥<br />
⎥<br />
⎥⎦<br />
K×<br />
P z k = e<br />
j 2πf<br />
D,kTPRI<br />
D = diag χ , χ , K,<br />
χ } , (3-12c)<br />
{ z<br />
2<br />
{<br />
1 2 K<br />
Φ = diag K<br />
. (3-12d)<br />
1,<br />
z , , z K<br />
} j 2πfD,kTPRI<br />
z k e<br />
Given these decomposed <strong>for</strong>ms, the matrix pencil A 1 –λA 0 can be expressed as:<br />
A<br />
1<br />
λA<br />
0<br />
= Z<br />
LD Φ − λI]<br />
=<br />
− [ Z . (3-13)<br />
From the Φ–λI term in the above expression, it can be seen that if λ equals to any of the main<br />
diagonal entries in Φ (i.e. any of z k ), then the rank of A 1 –λA 0 is reduced by one. Hence, in the<br />
absence of noise, values of the set {z k } are indeed the GEs of A 1 –λA 0 , and correspondingly the<br />
k th principal Doppler frequency f D,k can be found from the k th GE as follows:<br />
R
Chapter 3. Velocity Estimation in <strong>Color</strong> <strong>Flow</strong> Imaging: A Matrix Pencil Approach 67<br />
f<br />
1<br />
{ λ }<br />
D , k<br />
= arg k<br />
. (3-14)<br />
2πTPRI<br />
As reviewed by van der Veen et al. (1993), the solution <strong>for</strong> Φ in (3-13) is equivalent to finding a<br />
rotational matrix operator such that the rotation is invariant from a subspace perspective (i.e. the<br />
two data matrices A 1 and A 0 span the same subspace). Hence, the Matrix Pencil estimator can be<br />
considered as an estimation method that is based on rotational invariance principles.<br />
Low-Rank Matrix Reduction<br />
In the general case where white noise is present, the decomposed matrix <strong>for</strong>ms shown in<br />
(3-11) would still hold if low-rank matrix reduction is applied to A 1 and A 0 be<strong>for</strong>e finding the<br />
GEs. The low-rank reduction can generally be per<strong>for</strong>med by using singular value decomposition<br />
(SVD) to create A 1 and A 0 from only their K largest singular components (Hua and Sarkar 1990).<br />
As pointed out by van der Veen et al. (1993), such approach is equivalent to a total-least-squares<br />
<strong>for</strong>mulation of the estimation problem from a data fitting perspective. In terms of its<br />
implementation, an efficient way of realizing low-rank reduction is to first compute the rankreduced<br />
equivalent of a full Hankel data matrix (as defined in (2-12)) and then create the matrix<br />
pair from subsets of the rank-reduced matrix (Hua and Sarkar 1991). As originally derived using<br />
state-space estimation principles (Kung et al. 1983), this method is based on the following<br />
relation between the full data matrix A and its rank-reduced equivalent Ã:<br />
~<br />
A = A + Δ =<br />
K<br />
∑<br />
k = 1<br />
P<br />
∑<br />
H<br />
H<br />
σ u v + σ u v , (3-15)<br />
k<br />
k<br />
k<br />
k<br />
k = K + 1<br />
where Δ represents the noise perturbation of the full data matrix, while σ k , u k , and v k are<br />
respectively the singular value, left singular vector (of size N D –P+1), and right singular vector<br />
(of size P) <strong>for</strong> the k th SVD component. Since the matrix elements of A 1 and A 0 only differ by<br />
one row (see (3-10)), it follows that the rank-reduced matrices à 1 and à 0 can be expressed as:<br />
~<br />
A<br />
1<br />
K<br />
~<br />
= last ( N<br />
D<br />
− P)<br />
rows of A = ∑σ<br />
ku<br />
<strong>for</strong> u<br />
1, k<br />
=<br />
k = 1<br />
1, k<br />
k<br />
T<br />
[ u (1), u (2), K,<br />
u ( N − P)<br />
] ,<br />
k<br />
k<br />
v<br />
H<br />
k<br />
k<br />
k<br />
D<br />
(3-16a)<br />
~<br />
A<br />
0<br />
K<br />
~<br />
= first ( N<br />
D<br />
− P)<br />
rows of A = ∑σ<br />
ku<br />
<strong>for</strong> u<br />
0, k<br />
=<br />
k = 1<br />
0, k<br />
T<br />
[ u (0), u (1), K,<br />
u ( N − P −1)<br />
] ,<br />
k<br />
k<br />
v<br />
H<br />
k<br />
k<br />
D<br />
(3-16b)
Chapter 3. Velocity Estimation in <strong>Color</strong> <strong>Flow</strong> Imaging: A Matrix Pencil Approach 68<br />
where u 0,k and u 1,k are the first and last (N D –P) samples of the k th left singular vector. The<br />
efficacy of low-rank reduction improves in general when there is a large difference between the<br />
signal eigen-rank K and the matrix dimension parameter P. In fact, as shown by Hua and Sarkar<br />
(1990) <strong>for</strong> rank-one signals, better reduction results can be achieved when P lies between the<br />
range from N D /3 to N D /2.<br />
To further improve the estimation accuracy, an additional round of low-rank matrix<br />
reduction may be carried out through a joint SVD truncation of the reduced matrix pair [Ã 1 , Ã 0 ].<br />
Nevertheless, this extra reduction does not lead to substantial per<strong>for</strong>mance improvement because<br />
its noise perturbation characteristics are asymptotically equivalent to the initial round of<br />
reduction (Hua and Sarkar 1991). It is also worth noting that, if the signal is purely consisted of<br />
undamped sinusoids, low-rank reduction can be more accurately carried out by using a <strong>for</strong>wardbackward<br />
approach that applies SVD truncation to an augmented data matrix <strong>for</strong>med from the<br />
signal samples and their complex conjugates (del Rio and Sarkar 1996). However, we found that<br />
the <strong>for</strong>ward-backward approach is not significantly useful in color flow data processing since the<br />
intrinsic presence of transit-time broadening effects gives rise to damping characteristics in the<br />
Doppler signal.<br />
Numerical Computation<br />
To solve <strong>for</strong> the GEs of the matrix pair defined by (3-10) or (3-16), we can first rewrite<br />
the GE equation in (3-9) as a standard eigenvalue equation. This algebraic manipulation can be<br />
done by multiplying both sides of the GE equation with A H 0 and moving all matrix terms to one<br />
side of the equation. The resulting eigenvalue equation is then given by:<br />
H −1<br />
H<br />
+<br />
A A ) A A q = λ q ⇔ [ A A − I]<br />
q = 0 , (3-17)<br />
(<br />
0 0 0 1<br />
0 1<br />
λ<br />
where the ‘+’ superscript refers to a pseudo-inverse that can be computed from a matrix’s<br />
singular value components (Moon and Stirling 2000, Sec. 7.3). <strong>Based</strong> on this expression, it can<br />
be seen that the GEs of A 1 –λA 0 (or à 1 –λà 0 ) are essentially the eigenvalues of A + 0 A 1 (or à + 0 à 1 ).<br />
Hence, it is possible to use a standard eigenvalue solving algorithm such as QR factorization or<br />
power iterations to compute the GEs (Golub and van Loan 1996, Ch. 7 and 8). Once the GEs are<br />
found, the K principal Doppler frequencies can be calculated using (3-14). Subsequently, the<br />
frequency estimates can be classified as to whether they correspond to clutter or blood echoes.<br />
As already described in Section 3.2.2, one intuitive way to per<strong>for</strong>m this classification is to make
Chapter 3. Velocity Estimation in <strong>Color</strong> <strong>Flow</strong> Imaging: A Matrix Pencil Approach 69<br />
Table 3-2. Summary on the Flops Needed in the Matrix Pencil Estimator<br />
Step Flops Rationale<br />
1. Form Hankel matrix<br />
of size (N D –P+1)×P<br />
0 • No computations needed <strong>for</strong> data structuring<br />
2. Compute SVD of the<br />
Hankel matrix<br />
O[P 3 ] • Same derivation as Step 2 in Table 2-2<br />
3. Obtain the filtered<br />
data matrix Ã<br />
O[P 2 ]<br />
• O[(N D –P+1)P] flops needed to obtain à from<br />
Σ Κ σ k u k v k<br />
H<br />
• For P=N D /2, O[(N D –P+1)P] ≈ O[P 2 ]<br />
4. Form the matrix pair<br />
[Ã 1 , Ã 0 ]<br />
0 • No computations needed <strong>for</strong> data structuring<br />
5. Find the GEs of the<br />
matrix pencil à 1 –λà 0<br />
O[P 3 ]<br />
• An eigenvalue solver such as QR algorithm needs<br />
O[P 3 ] flops <strong>for</strong> the P×P matrix à 0 + à 1 (see Golub<br />
and Van Loan 1996, Ch. 7 and 8)<br />
use of the fact that blood echoes generally give rise to higher Doppler frequencies. As such, it is<br />
possible to identify the modal frequency of blood echoes from the principal spectral estimate that<br />
has the largest magnitude. This notion can be mathematically expressed as follows:<br />
{ f ∀ k ∈[1,<br />
]}<br />
f<br />
D(est)<br />
= f<br />
D, k<br />
<strong>for</strong> kb<br />
= arg max<br />
b D,<br />
k<br />
K . (3-18)<br />
In terms of its computational load, it can be expected that the Matrix Pencil estimator has<br />
a higher burden than the lag-one autocorrelator. In particular, the Matrix Pencil involves the use<br />
of SVD in order to carry out low-rank reduction on the matrix pair [A 1 , A 0 ], and it also requires<br />
the use of an eigenvalue solver to compute the principal frequency estimates. A breakdown of<br />
this estimator’s overall computational load is provided in Table 3-2, which shows the number of<br />
floating point operations (flops) needed <strong>for</strong> each step in the estimator. As can be noted, the<br />
Matrix Pencil in theory needs on the order of P 3 flops <strong>for</strong> each estimation run. There<strong>for</strong>e, this<br />
estimator is slightly more efficient than other advanced flow estimators that generally require on<br />
the order of N 3 D flops (Vaitkus et al. 1998), but it is two orders of magnitude more complex than<br />
the lag-one autocorrelator that only requires on the order of N D flops (by inspection of (3-2)). To<br />
improve the computational efficiency of the Matrix Pencil, it is possible to use more efficient<br />
SVD algorithms and eigenvalue solvers that exploit the Hankel structure inherent in the data<br />
matrices of this estimator (e.g. see Badeau et al. 2004 <strong>for</strong> a recent reference).<br />
k
Chapter 3. Velocity Estimation in <strong>Color</strong> <strong>Flow</strong> Imaging: A Matrix Pencil Approach 70<br />
Relationship to <strong>Eigen</strong>-Analysis<br />
The theoretical <strong>for</strong>mulation of the Matrix Pencil estimator can actually be linked to<br />
eigen-analysis. In particular, Matrix Pencil can be considered as a signal-domain <strong>for</strong>mulation of<br />
a modified eigen-based estimator that involves data smoothing. This relationship can be seen by<br />
first noting that the Hankel data matrix A in the Matrix Pencil can be converted into the<br />
following correlation matrix:<br />
R<br />
x<br />
≈ A<br />
T<br />
A<br />
∗<br />
⎡ R0<br />
(0)<br />
⎢<br />
R0<br />
(1)<br />
= ⎢<br />
⎢ M<br />
⎢<br />
⎣R0<br />
( P −1)<br />
R ( −1)<br />
R (0)<br />
R ( P − 2)<br />
1<br />
1<br />
1<br />
M<br />
L<br />
L<br />
O<br />
L<br />
RP−<br />
1(<br />
−P<br />
+ 1) ⎤<br />
RP<br />
1(<br />
P 2)<br />
⎥<br />
−<br />
− +<br />
⎥<br />
M ⎥<br />
⎥<br />
RP−<br />
1(0)<br />
⎦<br />
<strong>for</strong> R<br />
k<br />
( l)<br />
=<br />
P×<br />
P<br />
ND<br />
−P+<br />
max[0, l]<br />
+ k<br />
∑<br />
n=<br />
max[0, l]<br />
+ k<br />
x(<br />
n)<br />
x<br />
∗<br />
( n − l)<br />
,<br />
(3-19)<br />
where R k (l) is the smoothed autocorrelation estimate <strong>for</strong> the l th lag and is computed by summing<br />
the single-sample correlation values over a window of N D –P+1 samples (with k being the first<br />
sample index in the window when the lag is zero). From the properties of SVD (Moon and<br />
Stirling 2000, Sec. 7.1), it is well-known that the eigen-decomposition of this smoothed<br />
correlation matrix is related to the SVD of the Hankel data matrix as follows:<br />
A =<br />
P<br />
P<br />
H T ∗<br />
2 H<br />
∑σ ku<br />
k<br />
v<br />
k<br />
⇔ A A = ∑σ<br />
k<br />
v<br />
k<br />
v<br />
k<br />
, (3-20)<br />
k = 1<br />
k = 1<br />
where the eigenvalues and eigenvectors of A T A * are respectively the squared singular values and<br />
the right singular vectors of A. <strong>Based</strong> on such property and the principles of low-rank reduction,<br />
the Matrix Pencil can there<strong>for</strong>e be seen as being similar to an eigen-based estimator that involves<br />
the computation of GEs <strong>for</strong> a rank-reduced matrix pair <strong>for</strong>med from the K largest eigencomponents<br />
of the smoothed correlation matrix. Note that the eigen-decomposition equivalent<br />
of the Matrix Pencil was studied extensively by Stoica and Soderstrom (1991), and it is simply a<br />
data-smoothed version of the “estimation of signal parameters via rotational invariance<br />
techniques” (ESPRIT) framework originally proposed by Roy and Kailath (1989). Besides datasmoothed<br />
ESPRIT, it should be pointed out that a data-smoothed version of MUSIC also exists<br />
in practice (Stoica and Soderstrom 1991). In color flow data processing, this smoothed-MUSIC<br />
estimator is more useful than the original one described in Section 3.2.2 because multiple signal<br />
snapshots are generally not available to estimate the correlation matrix of each sample volume.
Chapter 3. Velocity Estimation in <strong>Color</strong> <strong>Flow</strong> Imaging: A Matrix Pencil Approach 71<br />
Table 3-3. Three Types of Matrix Pencil Estimators <strong>for</strong> <strong>Color</strong> <strong>Flow</strong> Data <strong>Processing</strong><br />
Type Principle Main Usage Limitation<br />
Rank-One<br />
(K = 1)<br />
Obtain one principal<br />
frequency estimate<br />
from a data ensemble<br />
Find mode frequency<br />
of filtered Doppler data<br />
Cannot be applied<br />
directly to raw Doppler<br />
data to estimate flow<br />
Rank-Two<br />
(K = 2)<br />
Extract two principal<br />
frequencies from a data<br />
ensemble<br />
Find modal flow<br />
estimates in presence of<br />
narrowband clutter<br />
Estimates may be biased<br />
when applied to data<br />
with wideband clutter<br />
Rank-Adaptive<br />
(K = K nom )<br />
Compute principal<br />
frequencies using an<br />
eigen-rank found from<br />
spectral spread analysis<br />
Find modal flow<br />
estimates in presence of<br />
any type of clutter<br />
Assume clutter spans a<br />
prescribed bandwidth<br />
3.3.3 Applications in <strong>Color</strong> <strong>Flow</strong> Data <strong>Processing</strong><br />
Rank-One Matrix Pencil<br />
As outlined in Table 3-3, the Matrix Pencil estimator has several potential applications in<br />
color flow data processing. For instance, by assuming a rank-one signal eigen-structure (i.e.<br />
setting K = 1), this parametric estimator can be used to compute the mode frequency of any postfilter<br />
Doppler signal. Note that the rank-one Matrix Pencil estimator can be considered as a<br />
generalized <strong>for</strong>m of the lag-one autocorrelator given in (3-2). Specifically, the rank-one<br />
estimator is a more advanced autocorrelator that has low-rank reduction capabilities depending<br />
on the choice of the pencil dimension P. As such, when the pencil dimension is greater than one,<br />
rank-one Matrix Pencil is less prone to noise perturbations than the original lag-one<br />
autocorrelator. It should be pointed out that the two estimators become equivalent in their <strong>for</strong>m<br />
when the pencil dimension is set equal to one (i.e. when P = 1). This equivalence can be shown<br />
by noting that, <strong>for</strong> a given post-filter Doppler signal y(n), the matrices A 1 and A 0 in the Matrix<br />
Pencil estimator degenerate to the following vectors (of size N D –1) in the limiting case:<br />
[ y( 1), y(2),<br />
K,<br />
y(<br />
−1<br />
] T<br />
a , (3-21a)<br />
1<br />
= N<br />
D<br />
)<br />
[ y( 0), y(1),<br />
K,<br />
y(<br />
− 2 ] T<br />
a . (3-21b)<br />
0<br />
= N<br />
D<br />
)<br />
From the solution of (3-17) and the expression in (3-14), a 1 and a 0 would <strong>for</strong>m a matrix pencil<br />
that gives the following modal frequency estimate:
Chapter 3. Velocity Estimation in <strong>Color</strong> <strong>Flow</strong> Imaging: A Matrix Pencil Approach 72<br />
H<br />
1<br />
a<br />
0<br />
a1<br />
f<br />
D<br />
= arg{}<br />
λ <strong>for</strong> λ = =<br />
H<br />
2πT<br />
a a<br />
PRI<br />
0<br />
0<br />
Ry<br />
(1)<br />
, (3-21c)<br />
R (0)<br />
y<br />
where R y (0) and R y (1) are respectively the lag-zero and lag-one autocorrelation functions of the<br />
filtered Doppler signal. Since R y (0) is always a real quantity (because it is simply the squarednorm<br />
of a 0 ), (3-21c) is essentially equivalent to (3-2), thus indicating that rank-one Matrix Pencil<br />
is the same as the lag-one autocorrelator when the pencil dimension is equal to one.<br />
Rank-Two Matrix Pencil<br />
Besides computing modal frequencies from filtered Doppler data, the Matrix Pencil can<br />
also be used to obtain flow estimates in the presence of clutter. In particular, by assuming a<br />
higher-rank signal eigen-structure, the Matrix Pencil can be applied directly to the raw Doppler<br />
data to extract the modal Doppler frequency of blood echoes. One type of higher-rank Matrix<br />
Pencil estimator that is useful in color flow data processing is the rank-two approach that models<br />
the raw Doppler signal as a sum of two principal complex sinusoids (i.e. it assumes K = 2). The<br />
rank-two Matrix Pencil, which is similar to second-order AR modeling and rank-two MUSIC,<br />
assumes that one of the principal complex sinusoids corresponds to clutter and the other<br />
corresponds to blood echoes. In turn, it obtains the modal frequency of blood echoes from the<br />
larger (magnitude-wise) of the two principal frequencies. In terms of its use in flow studies, this<br />
rank-two estimator is suitable <strong>for</strong> finding the modal frequency of blood echoes in studies where<br />
the clutter is narrowband in nature with respect to the Doppler spectral resolution. For instance,<br />
as recently demonstrated (Yu et al. 2005), it can be used to compute flow velocities from color<br />
flow data acquired from a flow phantom that produces stationary clutter.<br />
It should be pointed out that rank-two Matrix Pencil essentially degenerates to a secondorder<br />
AR estimator when the pencil dimension is set equal to two (i.e. when P = 2). Specifically,<br />
in this limiting case, the matrix pair [A 1 , A 0 ] <strong>for</strong> Matrix Pencil and the matrix A + 0 A 1 as seen in<br />
(3-17) would take on the following <strong>for</strong>ms (Hua and Sarkar 1990):<br />
⎡ x(1)<br />
x(2)<br />
⎤<br />
A<br />
⎢<br />
⎥<br />
1<br />
=<br />
⎢<br />
M M<br />
⎥<br />
,<br />
⎢⎣<br />
x(<br />
N − 2) ( −1)<br />
⎥<br />
D<br />
x N<br />
D ⎦<br />
⎡ x(0)<br />
x(1)<br />
⎤<br />
A<br />
⎢<br />
⎥<br />
0<br />
=<br />
⎢<br />
M M<br />
⎥<br />
, (3-22a)(3-22b)<br />
⎢⎣<br />
x(<br />
N − 3) ( − 2) ⎥<br />
D<br />
x N<br />
D ⎦<br />
⎡0<br />
c<br />
+<br />
1 ⎤<br />
A<br />
0<br />
A1<br />
= ⎢ ⎥ , (3-22c)<br />
⎣1<br />
c2<br />
⎦
Chapter 3. Velocity Estimation in <strong>Color</strong> <strong>Flow</strong> Imaging: A Matrix Pencil Approach 73<br />
where c 1 and c 2 are respectively the product coefficients between A + 0 and the last column of A 1 .<br />
As described in algebra textbooks (e.g. see Moon and Stirling 2000, Sec. 8.5.1), the matrix<br />
shown in (3-22c) is the same as the companion matrix of a second-order AR polynomial, and its<br />
two eigenvalues are indeed the roots of the AR polynomial. There<strong>for</strong>e, rank-two Matrix Pencil<br />
and second-order AR estimator are equivalent in principles when a pencil dimension of two is<br />
used. The theoretical advantage of using rank-two Matrix Pencil as opposed to second-order AR<br />
modeling is that the Matrix Pencil estimates can be made less sensitive to noise perturbations by<br />
per<strong>for</strong>ming low-rank reduction via the use of a larger pencil dimension (just like the rank-one<br />
<strong>for</strong>mulation). Note that, to achieve the same advantage with AR modeling, the original Prony’s<br />
method given in (3-6) may be modified to include a SVD truncation step be<strong>for</strong>e calculating the<br />
model coefficients (Tufts and Kumaresan 1982). Nevertheless, Hua and Sarkar (1990) have<br />
shown that the SVD-based Prony’s method is not as effective as the Matrix Pencil.<br />
Rank-Adaptive Matrix Pencil<br />
As already pointed out, the rank-two Matrix Pencil estimator inherently assumes that<br />
clutter in the Doppler signal can be adequately modeled as a single complex sinusoid. However,<br />
such assumption may not always be valid because tissue motion can give rise to clutter that is<br />
more wideband in nature. Consequently, rank-two Matrix Pencil may sometimes give modal<br />
flow estimates that are inconsistent with the actual flow dynamics. The inconsistency problem is<br />
generally more significant <strong>for</strong> Doppler data acquired using higher frequencies or longer<br />
ensemble periods because the Doppler spectral resolution is finer in these cases.<br />
To address the theoretical limitation of rank-two Matrix Pencil, it is worthwhile to<br />
develop an algorithm <strong>for</strong> Matrix Pencil to adaptively select its eigen-structure rank based on the<br />
Doppler spectral characteristics. An intuitive way of designing such algorithm is to analyze the<br />
spectral spread of the Matrix Pencil frequency estimates <strong>for</strong> different eigen-structure ranks. In<br />
particular, it can be expected that, <strong>for</strong> cases with dominating wideband clutter, the spectral<br />
spread of lower-rank Matrix Pencil estimates should be smaller because the most dominant<br />
frequencies would likely correspond to clutter. Hence, as illustrated in the flowchart in Fig. 3-4,<br />
the rank selection algorithm can involve a search <strong>for</strong> the minimum eigen-structure rank that<br />
yields Matrix Pencil estimates with spectral spread greater than a certain threshold Δf thr . The<br />
nominal rank K nom obtained from this algorithm can mathematically be expressed as follows:
Chapter 3. Velocity Estimation in <strong>Color</strong> <strong>Flow</strong> Imaging: A Matrix Pencil Approach 74<br />
Initialize Rank<br />
K = 2<br />
Get Rank-Reduced<br />
Data Matrix Ã<br />
Form Matrix Pencil<br />
à 1<br />
–λ Ã 0<br />
Find Generalized<br />
<strong>Eigen</strong>values of à 0+<br />
à 1<br />
Increment Rank<br />
K = K+1<br />
No<br />
Freq Spread<br />
> Δf thr<br />
?<br />
Yes<br />
Current Rank is<br />
the Nominal Rank<br />
Fig. 3-4. <strong>Flow</strong>chart of the rank selection algorithm used <strong>for</strong> rank-adaptive Matrix Pencil.<br />
The algorithm is based on analyzing the spectral spread of Matrix Pencil frequency estimates<br />
obtained <strong>for</strong> each eigen-structure rank.<br />
K<br />
( max{ f , f ,..., f } − min{ f , f f } > Δf<br />
)<br />
nom<br />
arg min<br />
D,1 D,2 D, K<br />
D,1 D,2<br />
,...,<br />
K<br />
= . (3-23)<br />
Note that the efficacy of this rank selection algorithm primarily depends on the choice of the<br />
spectral spread threshold Δf thr , which is a quantity analogous to the stopband of a clutter filter.<br />
Once the rank has been adaptively estimated, the modal frequency of blood echoes can then be<br />
set equal to the frequency with largest magnitude in the set of K nom Matrix Pencil estimates.<br />
It is worth pointing out that, in the signal processing field, there exist some other rank<br />
selection criteria such as the Akiake In<strong>for</strong>mation Criterion and the Minimal Description Length.<br />
These criteria generally work by finding the eigen-structure rank that gives the minimum fitting<br />
error with respect to the Doppler signal (Hayes 1996, Sec 8.5). However, they are not suitable<br />
<strong>for</strong> color flow data processing because parametric modeling is merely used in this application to<br />
obtain the modal flow estimate instead of finding the best model fit <strong>for</strong> the given Doppler data.<br />
Indeed, we recently showed that the Minimal Description Length criterion tends to overestimate<br />
the eigen-structure rank owing to the presence of transit-time broadening in the blood component<br />
of Doppler signals (Yu and Cobbold 2006).<br />
D, K<br />
thr
Chapter 3. Velocity Estimation in <strong>Color</strong> <strong>Flow</strong> Imaging: A Matrix Pencil Approach 75<br />
3.4 Simulation <strong>Methods</strong><br />
3.4.1 Method of Study<br />
General Overview<br />
To examine the theoretical per<strong>for</strong>mance of the Matrix Pencil estimation framework, we<br />
made use of the simulation model that was previously developed to analyze the Hankel-SVD<br />
filter (see Section 2.4.1). In particular, two series of simulations were conducted to<br />
quantitatively analyze the efficacy of Matrix Pencil under various flow scenarios and noise<br />
levels. The first series of simulations, which involves the modeling of a flow scenario with no<br />
tissue motion, aims to evaluate the per<strong>for</strong>mance of rank-one and rank-two Matrix Pencil<br />
estimators at different BSNRs and flow velocities as well as to demonstrate these estimators’<br />
low-rank reduction capabilities <strong>for</strong> various pencil dimensions. As a comparative assessment, we<br />
also studied the estimation per<strong>for</strong>mance of other frequency-based velocity estimators (as<br />
discussed in Section 3.2) in this flow scenario. With the insights gained from the initial<br />
simulations, a second series of simulations based on a flow scenario with moving tissue was then<br />
per<strong>for</strong>med to study the efficacy of rank-adaptive Matrix Pencil and the spectral spread criterion<br />
proposed <strong>for</strong> rank selection. The per<strong>for</strong>mance analysis was also carried out with rank-one and<br />
rank-two Matrix Pencil estimators to gain comparative insights on the advantages of adaptive<br />
rank selection.<br />
Per<strong>for</strong>mance Measures<br />
In this simulation study, both first-order and second-order statistical measures were used<br />
to quantitatively assess Matrix Pencil’s estimation per<strong>for</strong>mance. In particular, the two<br />
per<strong>for</strong>mance measures that were considered are: 1) mean estimation bias, and 2) root-meansquared<br />
(RMS) estimation error. The mean estimation bias was computed by finding the average<br />
difference between the estimated velocity and the actual blood velocity used to synthesize the<br />
Doppler signal. It is well-established from estimation theory that a low bias is an indication of<br />
high estimation accuracy. On the other hand, the RMS estimation error was calculated by taking<br />
the square root of the average squared difference between the estimated velocity and the true<br />
blood velocity. For an estimator to have high accuracy and precision, its RMS estimation error<br />
should be low in general.
Chapter 3. Velocity Estimation in <strong>Color</strong> <strong>Flow</strong> Imaging: A Matrix Pencil Approach 76<br />
Table 3-4. Simulation Parameters Used in the Estimator Analysis<br />
Parameter<br />
Value<br />
Fixed Parameters<br />
<strong>Ultrasound</strong> propagation speed, c o<br />
1540 m/s<br />
F-number, F/W 4<br />
Pulse carrier frequency, f o<br />
5 MHz<br />
Pulse repetition interval, T PRI<br />
1.0 ms<br />
Beam-flow angle, θ 60°<br />
Doppler ensemble size, N D 10<br />
Dynamic range<br />
14 bits<br />
Number of realizations per dataset 10000<br />
Clutter-to-blood signal ratio<br />
30 dB<br />
Average noise strength, κ n<br />
10 dB<br />
Clutter vibration frequency, f vib<br />
5 Hz<br />
Variable Parameters<br />
Blood velocity, v b<br />
0 to v alias<br />
Blood-signal-to-noise ratio, BSNR -20 to +30 dB<br />
Maximum tissue velocity, v c,max<br />
0 or 1 cm/s<br />
Responding Parameters<br />
Aliasing velocity, v alias<br />
15.4 cm/s<br />
Clutter phase modulation index, φ c,max 0 or 13.0 radians<br />
Maximum clutter frequency<br />
0 or 70 Hz<br />
3.4.2 Simulation Parameters<br />
Data Synthesis<br />
The data synthesis parameters used in the simulations are listed in Table 3-4. For this<br />
study, blood velocity (v b ) and blood-signal-to-noise ratio (BSNR) are the two major variable<br />
parameters of interest. Specifically, color flow datasets (each with 10000 realizations) were<br />
synthesized <strong>for</strong> blood velocities ranging from zero to the aliasing limit (15.4 cm/s) as well as<br />
BSNRs ranging from -20 to 30 dB in order to assess the per<strong>for</strong>mance of the Matrix Pencil<br />
estimator. A third parameter that was varied in this study is the maximum clutter velocity<br />
(v c,max ). This parameter was either set to 0 or 1 cm/s to simulate narrowband Doppler clutter as<br />
well as wideband ones. In terms of the major fixed parameters, the ultrasound frequency (f o ) and<br />
the pulse repetition interval (T PRI ) were respectively set to 5 MHz and 1.0 ms in order to model<br />
data acquisition in a typical low-velocity vascular imaging scenario. As well, the average<br />
strength of Doppler clutter was defined to be 30 dB greater than that of blood echoes, and a<br />
Doppler ensemble size of 10 samples was used <strong>for</strong> the synthesized datasets.
Chapter 3. Velocity Estimation in <strong>Color</strong> <strong>Flow</strong> Imaging: A Matrix Pencil Approach 77<br />
Data Analysis<br />
For this study, flow estimation was per<strong>for</strong>med using four different frequency-based<br />
methods: lag-one autocorrelation, AR modeling, Matrix Pencil, and data-smoothed Root-<br />
MUSIC. The estimation process was carried out on all the synthesized Doppler datasets to<br />
assess the mean bias and the RMS error of the four estimators. Note that <strong>for</strong> the autocorrelator<br />
and the rank-one Matrix Pencil estimator, a fifth-order, projection-initialized IIR filter with a 100<br />
Hz nominal cutoff was used to suppress Doppler clutter prior to flow estimation † . As <strong>for</strong> the<br />
rank-adaptive Matrix Pencil estimator, a two-way spectral spread threshold of 150 Hz was used<br />
<strong>for</strong> the rank selection algorithm (i.e. Δf thr = 150 Hz in (3-23)). It should also be pointed out that,<br />
to avoid aliasing problems during the analysis, all flow estimates were unwrapped in velocity so<br />
that they are bounded within the range v b ±v alias , where v alias is the aliasing velocity.<br />
3.5 Simulation Results<br />
3.5.1 <strong>Flow</strong> Scenario with Stationary Clutter<br />
Per<strong>for</strong>mance of Low-Rank Reduction<br />
As an assessment of the Matrix Pencil’s low-rank reduction capabilities, Fig. 3-5 shows<br />
the bias and the RMS error of this estimator <strong>for</strong> different pencil dimensions as a function of<br />
BSNR. Results are provided <strong>for</strong> both rank-one and rank-two Matrix Pencil when the Doppler<br />
ensemble size is 10 samples, and they were obtained by processing Doppler data synthesized<br />
with stationary clutter and a 5 cm/s blood velocity. As can be seen, the estimation bias of both<br />
rank-one and rank-two Matrix Pencil drops significantly at medium and low BSNRs as the pencil<br />
dimension increases. In addition, the RMS error appears to decrease gradually when the pencil<br />
dimension becomes larger. Both of these observations show that low-rank reduction allows<br />
Matrix Pencil to have better resilience against background white noise. They also suggest that<br />
the flow estimation per<strong>for</strong>mance is generally better at larger pencil dimensions: a result that can<br />
be expected because larger pencil dimensions give rise to more entries in the Hankel data matrix<br />
and in turn the rank reduction per<strong>for</strong>mance should be improved. <strong>Based</strong> on these grounds, it<br />
appears that the dimension parameter <strong>for</strong> the Matrix Pencil estimator should be set to the<br />
maximum possible value (i.e. N D /2).<br />
† Other types of filters may be used instead to obtain similar per<strong>for</strong>mance insights <strong>for</strong> these two estimators.
Chapter 3. Velocity Estimation in <strong>Color</strong> <strong>Flow</strong> Imaging: A Matrix Pencil Approach 78<br />
Estimation Bias [cm/s]<br />
(a) Bias (Rank-One Matrix Pencil)<br />
14<br />
12<br />
10<br />
8<br />
6<br />
4<br />
2<br />
P=1<br />
P=2<br />
P=3<br />
P=4<br />
P=5<br />
0<br />
-20 -15 -10 -5 0 5 10 15 20 25 30<br />
Blood-<strong>Signal</strong>-to-Noise Ratio [dB]<br />
RMS Error [cm/s]<br />
(b) RMS Error (Rank-One Estimator)<br />
14<br />
12<br />
10<br />
8<br />
6<br />
4<br />
2<br />
P=1<br />
P=2<br />
P=3<br />
P=4<br />
P=5<br />
0<br />
-20 -15 -10 -5 0 5 10 15 20 25 30<br />
Blood-<strong>Signal</strong>-to-Noise Ratio [dB]<br />
Estimation Bias [cm/s]<br />
14<br />
12<br />
10<br />
8<br />
6<br />
4<br />
2<br />
(c) Bias (Rank-Two Estimator)<br />
P=2<br />
P=3<br />
P=4<br />
P=5<br />
0<br />
-20 -15 -10 -5 0 5 10 15 20 25 30<br />
Blood-<strong>Signal</strong>-to-Noise Ratio [dB]<br />
RMS Error [cm/s]<br />
(d) RMS Error (Rank-Two Estimator)<br />
14<br />
12<br />
10<br />
8<br />
6<br />
4<br />
2<br />
P=2<br />
P=3<br />
P=4<br />
P=5<br />
0<br />
-20 -15 -10 -5 0 5 10 15 20 25 30<br />
Blood-<strong>Signal</strong>-to-Noise Ratio [dB]<br />
Fig. 3-5. Estimation per<strong>for</strong>mance of rank-one and rank-two Matrix Pencil estimators in the<br />
flow scenario with stationary clutter. In (a) and (b), the estimation biases and the RMS errors<br />
are shown <strong>for</strong> the rank-one estimator at various BSNRs and pencil dimensions. The<br />
corresponding results <strong>for</strong> the rank-two estimator are shown in (c) and (d). Note that the true<br />
blood velocity used to synthesize the Doppler data is 5 cm/s.<br />
Comparative Assessment at Different BSNRs<br />
In Fig. 3-6, the estimation per<strong>for</strong>mance of Matrix Pencil at various BSNRs is compared<br />
against the autocorrelator, the second-order AR estimator, and the rank-two Root-MUSIC<br />
estimator (with data smoothing). The results shown are based on Doppler data synthesized with<br />
actual blood velocities of 5 and 10 cm/s (i.e. mean blood Doppler frequencies of 162 and 325<br />
Hz). In other words, these results respectively correspond to cases where the blood frequency
Chapter 3. Velocity Estimation in <strong>Color</strong> <strong>Flow</strong> Imaging: A Matrix Pencil Approach 79<br />
Estimation Bias [cm/s]<br />
14<br />
12<br />
10<br />
8<br />
6<br />
4<br />
2<br />
(a) Bias (v b = 5 cm/s)<br />
ACR<br />
AR<br />
MUSIC<br />
RK-1 MP<br />
RK-2 MP<br />
0<br />
-20 -15 -10 -5 0 5 10 15 20 25 30<br />
Blood-<strong>Signal</strong>-to-Noise Ratio [dB]<br />
RMS Error [cm/s]<br />
14<br />
12<br />
10<br />
8<br />
6<br />
4<br />
2<br />
(b) RMS Error (v b = 5 cm/s)<br />
ACR<br />
AR<br />
MUSIC<br />
RK-1 MP<br />
RK-2 MP<br />
0<br />
-20 -15 -10 -5 0 5 10 15 20 25 30<br />
Blood-<strong>Signal</strong>-to-Noise Ratio [dB]<br />
Estimation Bias [cm/s]<br />
14<br />
12<br />
10<br />
8<br />
6<br />
4<br />
2<br />
(c) Bias (v b = 10 cm/s)<br />
ACR<br />
AR<br />
MUSIC<br />
RK-1 MP<br />
RK-2 MP<br />
0<br />
-20 -15 -10 -5 0 5 10 15 20 25 30<br />
Blood-<strong>Signal</strong>-to-Noise Ratio [dB]<br />
RMS Error [cm/s]<br />
14<br />
12<br />
10<br />
8<br />
6<br />
4<br />
2<br />
(d) RMS Error (v b = 10 cm/s)<br />
ACR<br />
AR<br />
MUSIC<br />
RK-1 MP<br />
RK-2 MP<br />
0<br />
-20 -15 -10 -5 0 5 10 15 20 25 30<br />
Blood-<strong>Signal</strong>-to-Noise Ratio [dB]<br />
Fig. 3-6. Estimation per<strong>for</strong>mance of various frequency-based flow estimators in the flow<br />
scenario with stationary clutter. In (a) and (b), the estimation bias and the RMS error at<br />
various BSNRs are shown <strong>for</strong> the case where the actual blood velocity is 5 cm/s. The<br />
corresponding results <strong>for</strong> the 10 cm/s blood velocity case are shown in (c) and (d).<br />
contents are near and away from the stopband of the clutter filter. Note that, when using the<br />
Matrix Pencil estimator to process the synthesized Doppler data, the pencil dimension was<br />
defined as N D /2 to maximize the low-rank reduction per<strong>for</strong>mance. Likewise, when using the<br />
data-smoothed MUSIC estimator, a window size of N D /2 was chosen to <strong>for</strong>m the Doppler<br />
correlation matrix.<br />
From the results shown, it can be seen that rank-one Matrix Pencil (black dashed line) is<br />
less biased than the autocorrelator (gray dashed line) at any BSNR. The rank-one Matrix
Chapter 3. Velocity Estimation in <strong>Color</strong> <strong>Flow</strong> Imaging: A Matrix Pencil Approach 80<br />
Estimation Bias [cm/s]<br />
8<br />
6<br />
4<br />
(a) Bias (BSNR = 10 dB)<br />
ACR<br />
AR<br />
MUSIC<br />
RK-1 MP<br />
RK-2 MP<br />
2<br />
0<br />
-2<br />
-4<br />
0 2 4 6 8 10 12 14<br />
Actual Blood Velocity [cm/s]<br />
RMS Error [cm/s]<br />
14<br />
12<br />
10<br />
(b) RMS Error (BSNR = 10 dB)<br />
8<br />
6<br />
4<br />
2<br />
ACR<br />
AR<br />
MUSIC<br />
RK-1 MP<br />
RK-2 MP<br />
0<br />
0 2 4 6 8 10 12 14<br />
Actual Blood Velocity [cm/s]<br />
Estimation Bias [cm/s]<br />
8<br />
6<br />
4<br />
2<br />
0<br />
-2<br />
-4<br />
(c) Bias (BSNR = 0 dB)<br />
ACR<br />
AR<br />
MUSIC<br />
RK-1 MP<br />
RK-2 MP<br />
0 2 4 6 8 10 12 14<br />
Actual Blood Velocity [cm/s]<br />
RMS Error [cm/s]<br />
14<br />
12<br />
10<br />
8<br />
6<br />
4<br />
2<br />
(d) RMS Error (BSNR = 0 dB)<br />
ACR<br />
AR<br />
MUSIC<br />
RK-1 MP<br />
RK-2 MP<br />
0<br />
0 2 4 6 8 10 12 14<br />
Actual Blood Velocity [cm/s]<br />
Fig. 3-7. Estimator per<strong>for</strong>mance as a function of blood velocity in the flow scenario with<br />
stationary clutter. In (a) and (b), the estimation bias and the RMS error are shown <strong>for</strong> the<br />
case where the BSNR is 10 dB. In (c) and (d), similar results are shown <strong>for</strong> a BSNR of 0 dB.<br />
Pencil also appears to be more precise than the autocorrelator when the blood frequency contents<br />
are close to filter stopband (as seen <strong>for</strong> the case where the blood velocity is 5 cm/s). Both of<br />
these findings are consistent with our previous observations on the advantages of low-rank<br />
reduction in the Matrix Pencil estimator. Another observation to be noted is that, <strong>for</strong> the 5 cm/s<br />
blood velocity case, rank-one Matrix Pencil and the autocorrelator both remain biased at high<br />
BSNRs because of distortions from the clutter filter. On the other hand, all three rank-two flow<br />
estimators are not prone to these distortions since they do not involve clutter filtering. In<br />
comparison between the rank-two estimators, rank-two Matrix Pencil (black solid line) is
Chapter 3. Velocity Estimation in <strong>Color</strong> <strong>Flow</strong> Imaging: A Matrix Pencil Approach 81<br />
significantly more accurate than the second-order AR estimator (gray dotted line) at medium and<br />
low BSNRs, and it appears to share similar per<strong>for</strong>mance trends with the rank-two Root-MUSIC<br />
estimator (gray solid line). This result can be expected because Matrix Pencil and Root-MUSIC<br />
both involve an orthogonal decomposition (either through SVD or eigen-decomposition) in their<br />
theoretical <strong>for</strong>mulation to help separate signal components from noise (as noted in Section 3.3.2).<br />
Comparative Assessment at Different Blood Velocities<br />
As a generalization of the above observations, Fig. 3-7 shows the bias and the RMS error<br />
of various frequency-based flow estimators as a function of the true blood velocity. These<br />
results were found by processing Doppler data generated with BSNRs of 0 and 10 dB. As can be<br />
seen, the rank-one Matrix Pencil and the lag-one autocorrelator both exhibit substantial<br />
estimation bias at low blood velocities because of distortions from the clutter filter stopband. In<br />
fact, their bias is at a maximum <strong>for</strong> blood velocities near 4-5 cm/s (i.e. <strong>for</strong> blood Doppler<br />
frequencies around 130-160 Hz) since the blood spectral components in this case are mainly<br />
located inside the clutter filter’s transition region where signal distortions are more substantial.<br />
On the other hand, rank-two Matrix Pencil and rank-two Root-MUSIC both have significantly<br />
lower biases than the other estimators at any blood velocity. This observation is in accordance<br />
with our previous results that demonstrate the advantage of applying low-rank reduction to flow<br />
estimation. It is also worth pointing out that, at higher blood velocities, the precision of rank-one<br />
Matrix Pencil and the autocorrelator appears to be slightly better. This phenomenon can be<br />
explained by recognizing that, <strong>for</strong> these two estimators, the clutter filtering step prior to flow<br />
estimation has more or less suppressed some of the background white noise and hence the<br />
estimation variance is inherently lower.<br />
3.5.2 <strong>Flow</strong> Scenario with Non-Stationary Clutter<br />
Efficacy of Adaptive Rank Selection<br />
To demonstrate the need <strong>for</strong> adaptive rank selection when using Matrix Pencil, we first<br />
studied the estimation per<strong>for</strong>mance of rank-two Matrix Pencil isn the flow scenario with clutter<br />
generated by moving tissue. In this study, the flow estimates of rank-two Matrix Pencil were<br />
evaluated <strong>for</strong> cases with blood velocities of 5 and 10 cm/s as well as a BSNR of 10 dB. The<br />
upper half of Fig. 3-8 shows the corresponding distribution <strong>for</strong> 10000 estimates of the blood
Chapter 3. Velocity Estimation in <strong>Color</strong> <strong>Flow</strong> Imaging: A Matrix Pencil Approach 82<br />
(a) Rank-Two Matrix Pencil<br />
(v b = 5 cm/s)<br />
(b) Rank-Two Matrix Pencil<br />
(v b = 10 cm/s)<br />
Num. Estimates [1000's]<br />
1.4<br />
1.2<br />
1<br />
Clutter<br />
Range<br />
0.8<br />
0.6<br />
0.4<br />
0.2<br />
0<br />
-338 0 162 662<br />
Estimated Doppler Frequency [Hz]<br />
Num. Estimates [1000's]<br />
1.4<br />
1.2<br />
1<br />
Clutter<br />
Range<br />
0.8<br />
0.6<br />
0.4<br />
0.2<br />
0<br />
-175 0 325 825<br />
Estimated Doppler Frequency [Hz]<br />
(c) Rank-Adaptive Matrix Pencil<br />
(v b = 5 cm/s)<br />
(d) Rank-Adaptive Matrix Pencil<br />
(v b = 10 cm/s)<br />
Num. Estimates [1000's]<br />
1.4<br />
1.2<br />
1<br />
Clutter<br />
Range<br />
0.8<br />
0.6<br />
0.4<br />
0.2<br />
0<br />
-338 0 162 662<br />
Estimated Doppler Frequency [Hz]<br />
Num. Estimates [1000's]<br />
1.4<br />
1.2<br />
1<br />
Clutter<br />
Range<br />
0.8<br />
0.6<br />
0.4<br />
0.2<br />
0<br />
-175 0 325 825<br />
Estimated Doppler Frequency [Hz]<br />
Fig. 3-8. Histograms of flow frequency estimates obtained from Matrix Pencil. The results<br />
<strong>for</strong> the rank-two estimator are shown in (a) and (b), while the ones <strong>for</strong> the rank-adaptive<br />
estimator are shown in (c) and (d). These histograms correspond to cases where the actual<br />
blood velocity is 5 or 10 cm/s. Note that the flow estimates are displayed over the unaliased<br />
Doppler spectral range, and 10000 estimates were used to <strong>for</strong>m each histogram.<br />
Doppler frequency (prior to velocity conversion) in the two cases. As can be seen, a number of<br />
rank-two Matrix Pencil estimates are biased towards zero: a result that is inconsistent with the<br />
flow dynamics since blood velocities of 5 and 10 cm/s respectively give rise to mean blood<br />
Doppler frequencies of 162 and 325 Hz. In fact, when the blood velocity is 5 cm/s, it was found<br />
that 19.2% of the flow estimates in this dataset have a magnitude lower than the approximate<br />
clutter range of ±70 Hz; likewise, when the blood velocity is 10 cm/s, 13.9% of the estimates<br />
were less than the peak clutter frequency. In contrast to the rank-two estimation results, the<br />
bottom half of Fig. 3-8 shows the distribution of blood Doppler frequencies found from a rankadaptive<br />
Matrix Pencil estimator that uses a spectral spread threshold of 150 Hz during rank
Chapter 3. Velocity Estimation in <strong>Color</strong> <strong>Flow</strong> Imaging: A Matrix Pencil Approach 83<br />
Estimation Bias [cm/s]<br />
Estimation Bias [cm/s]<br />
8<br />
6<br />
4<br />
2<br />
0<br />
-2<br />
-4<br />
(a) Bias (BSNR = 10 dB)<br />
ACR<br />
RK-1 MP<br />
RK-2 MP<br />
RK-A MP<br />
0 2 4 6 8 10 12 14<br />
Actual Blood Velocity [cm/s]<br />
8<br />
6<br />
4<br />
2<br />
0<br />
-2<br />
-4<br />
(c) Bias (BSNR = 0 dB)<br />
ACR<br />
RK-1 MP<br />
RK-2 MP<br />
RK-A MP<br />
0 2 4 6 8 10 12 14<br />
Actual Blood Velocity [cm/s]<br />
RMS Error [cm/s]<br />
RMS Error [cm/s]<br />
14<br />
12<br />
10<br />
(b) RMS Error (BSNR = 10 dB)<br />
8<br />
6<br />
4<br />
2<br />
ACR<br />
RK-1 MP<br />
RK-2 MP<br />
RK-A MP<br />
0<br />
0 2 4 6 8 10 12 14<br />
Actual Blood Velocity [cm/s]<br />
14<br />
12<br />
10<br />
8<br />
6<br />
4<br />
2<br />
(d) RMS Error (BSNR = 0 dB)<br />
ACR<br />
RK-1 MP<br />
RK-2 MP<br />
RK-A MP<br />
0<br />
0 2 4 6 8 10 12 14<br />
Actual Blood Velocity [cm/s]<br />
Fig. 3-9. Estimation per<strong>for</strong>mance of flow estimators in the flow scenario with non-stationary<br />
clutter. In (a) and (b), the estimation bias and the RMS error as a function of blood velocity<br />
are shown <strong>for</strong> the case where the BSNR is 10 dB. Similar results <strong>for</strong> a BSNR of 0 dB are<br />
shown in (c) and (d).<br />
selection. Although favorable results were not observed (some overestimation were seen) <strong>for</strong> the<br />
case with 5 cm/s velocity, the histogram <strong>for</strong> the higher velocity case (10 cm/s) revealed that<br />
rank-adaptive Matrix Pencil is able to obtain flow estimates that are relatively centered at the<br />
mean blood frequency of the synthesized Doppler data. As such, rank-adaptive Matrix Pencil<br />
appears to have better estimation accuracy than the rank-two counterpart at high blood velocities.<br />
Comparative Assessment at Different Blood Velocities<br />
As an assessment of the general per<strong>for</strong>mance of rank-adaptive Matrix Pencil, Fig. 3-9<br />
shows the bias and the RMS error of this estimator (solid line) <strong>for</strong> various blood velocities and
Chapter 3. Velocity Estimation in <strong>Color</strong> <strong>Flow</strong> Imaging: A Matrix Pencil Approach 84<br />
two BSNR values. For comparison, the per<strong>for</strong>mance curves are also plotted <strong>for</strong> the<br />
autocorrelator (dash-dotted line) as well as the rank-one (dotted line) and rank-two Matrix Pencil<br />
estimators (dashed line). From these figures, it can be seen that in the higher blood velocity<br />
range (> ½v alias ), rank-adaptive Matrix Pencil is the most accurate (i.e. lowest bias) amongst the<br />
four estimators. However, in the lower blood velocity range (< ½v alias ), rank-adaptive Matrix<br />
Pencil appears to suffer a significant drop in the estimation precision. Such result can be<br />
expected though because the rank selection algorithm used in this study assumes that the Doppler<br />
clutter spans a prescribed bandwidth and there<strong>for</strong>e tends to exhibit per<strong>for</strong>mance characteristics<br />
similar to the stopband of a clutter filter. In fact, as can be seen from Fig. 3-9, the estimation<br />
precision of rank-adaptive Matrix Pencil is quite similar to that <strong>for</strong> the autocorrelator and rankone<br />
Matrix Pencil (both of which involves the use of a clutter filter). It is also worth pointing out<br />
that the better accuracy and precision seen <strong>for</strong> rank-two Matrix Pencil at lower velocities is a<br />
somewhat misleading result since this estimator gives flow estimates that are always biased<br />
towards zero and hence it tends to yield lower estimation errors when the actual blood velocity is<br />
low. Note that the use of rank-two Matrix Pencil in flow scenarios with non-stationary clutter<br />
can be considered as being analogous to applying a correlation-based estimator to color flow data<br />
whose clutter has not been adequately suppressed.<br />
3.6 Concluding Remarks<br />
It has been the intent of this chapter to investigate the use of the Matrix Pencil estimation<br />
framework in color flow data processing. In our theoretical <strong>for</strong>mulation, we have first<br />
considered a type of fixed-rank Matrix Pencil estimator called rank-one Matrix Pencil and have<br />
shown that it is a generalized <strong>for</strong>m of the lag-one autocorrelator that is often used in color flow<br />
data processing. To realize the potential of Matrix Pencil in obtaining flow estimates directly<br />
from raw color flow data (i.e. unfiltered Doppler data), we have considered the use of a rank-two<br />
Matrix Pencil estimator and have shown that it is suitable <strong>for</strong> use in flow scenarios where the<br />
Doppler clutter can be sufficiently modeled as a single complex sinusoid. Besides fixed-rank<br />
Matrix Pencil estimators, we have also developed an adaptive-rank Matrix Pencil estimator that<br />
can be used to obtain flow estimates in the presence of wideband Doppler clutter. The rank<br />
selection algorithm used in rank-adaptive Matrix Pencil is based on a search of the minimum<br />
eigen-structure rank that yields a Matrix Pencil spectral spread greater than a certain bandwidth.
Chapter 3. Velocity Estimation in <strong>Color</strong> <strong>Flow</strong> Imaging: A Matrix Pencil Approach 85<br />
Our simulation study has demonstrated that rank-one Matrix Pencil is more accurate and<br />
precise than the lag-one autocorrelator at various blood-signal-to-noise ratios. Results have also<br />
shown that, in flow scenarios with stationary Doppler clutter, rank-two Matrix Pencil is even<br />
more accurate than rank-one Matrix Pencil since the <strong>for</strong>mer does not suffer from potential biases<br />
due to clutter filtering. On the other hand, in flow scenarios with non-stationary Doppler clutter,<br />
rank-adaptive Matrix Pencil has shown to be more useful as it can give velocity estimates that<br />
are more consistent with the actual flow dynamics (provided that the blood spectral components<br />
are outside the specified clutter bandwidth). From these findings, it can be seen that the Matrix<br />
Pencil framework is potentially useful in color flow data processing.
CHAPTER 4<br />
Single-Module Approach to <strong>Color</strong> <strong>Flow</strong> Data<br />
<strong>Processing</strong>: An Experimental Study<br />
4.1 Chapter Overview<br />
In the previous two chapters, we considered how clutter suppression and flow estimation<br />
in color flow imaging can be per<strong>for</strong>med by making use of two algebraic quantities known as the<br />
Hankel matrix and the matrix pencil. Specifically, Chapter 2 has described an adaptive clutter<br />
filtering design that is based on principal Hankel component analysis of each Doppler signal<br />
vector. As well, Chapter 3 has presented a parametric flow estimation framework that treats<br />
Doppler spectral estimation as a generalized eigenvalue (GE) problem. In this chapter, the<br />
concepts presented in Chapters 2 and 3 are combined to <strong>for</strong>m a color flow data processor that<br />
per<strong>for</strong>ms flow detection and flow estimation within the same processing module. The proposed<br />
processor is then applied to a few different in vitro and in vivo flow imaging studies to gain<br />
insights on its flow estimation accuracy and flow detection per<strong>for</strong>mance.<br />
The remainder of this chapter is aimed at providing a system-level description of the<br />
proposed color flow data processor and presenting this processor’s efficacy on experimentally<br />
acquired color flow imaging data. To <strong>for</strong>mulate discussion, the chapter contents have been<br />
organized as follows:<br />
• Section 4.2 highlights the theoretical principles used by the new eigen-based color flow data<br />
processor and provides a conceptual overview of this design;<br />
• Section 4.3 presents results on the per<strong>for</strong>mance of the new color flow signal processor based<br />
on color flow imaging data acquired from an in vitro scenario;<br />
• Section 4.4 provides further results on the efficacy of the new color flow signal processor<br />
based on in vivo color flow imaging data acquired from two different cases;<br />
• Section 4.5 gives some additional comments on the proposed signal processor and<br />
summarizes the conclusions drawn from the experimental studies.<br />
- 86 -
Chapter 4. Single-Module Approach to <strong>Color</strong> <strong>Flow</strong> Data <strong>Processing</strong> 87<br />
(a) Conventional Processors<br />
Raw<br />
Doppler Data<br />
(b) Proposed Processor<br />
Raw<br />
Doppler Data<br />
Filter<br />
Parameters<br />
CLUTTER FILTER /<br />
FLOW DETECTION<br />
Filtered<br />
Doppler Data<br />
Clutter<br />
BW Thr.<br />
HANKEL-SVD +<br />
MATRIX PENCIL<br />
FLOW<br />
ESTIMATION<br />
Doppler<br />
Power<br />
<strong>Flow</strong><br />
Velocity<br />
Doppler<br />
Power<br />
<strong>Flow</strong><br />
Velocity<br />
Fig. 4-1. Block diagrams illustrating the difference between (a) conventional color flow<br />
data processors and (b) the proposed processor. Note that the proposed processor does not<br />
involve direct computation of the filtered Doppler data since flow detection and flow<br />
estimation are jointly per<strong>for</strong>med within the same module.<br />
4.2 The Single-Module <strong>Color</strong> <strong>Flow</strong> Data Processor<br />
4.2.1 Background Considerations<br />
Design Motivation<br />
As shown in Fig. 4-1a, color flow signal processing is conventionally per<strong>for</strong>med using a<br />
two-module approach that treats clutter filtering and flow estimation as independent processing<br />
modules. During operation, the clutter filtering module first attempts to distinguish the blood<br />
component of Doppler signal from low-frequency tissue clutter, and thus it can be considered as<br />
a flow detection module. From the filtered Doppler signal, the flow estimation module then<br />
finds the flow velocity and the flow signal power as quantitative indicators <strong>for</strong> the presence or<br />
absence of blood flow. The estimation per<strong>for</strong>mance of this module, however, much depends on<br />
the ability of the clutter filtering module to adequately suppress Doppler clutter without<br />
distorting blood echoes. Because of such dependence, the clutter filtering module is often<br />
regarded from a signal processing perspective as a principal source of error in the estimation of<br />
flow velocity and flow power. In attempt to avoid this problem associated with the two-module
Chapter 4. Single-Module Approach to <strong>Color</strong> <strong>Flow</strong> Data <strong>Processing</strong> 88<br />
color flow signal processing approach, we pursued the development of a new single-module<br />
signal processor that is based on a joint detection-estimation <strong>for</strong>mulation.<br />
Fundamental Principles<br />
Be<strong>for</strong>e the details of the new color flow data processor are considered, it is worth<br />
reviewing the basic principles of the Hankel-SVD filter and the Matrix Pencil estimator, which<br />
are the two signal processing frameworks presented in the previous chapters. First, as described<br />
in Chapter 2, the Hankel-SVD filter is an adaptive clutter suppression strategy based on the<br />
singular value decomposition (SVD) of a Hankel matrix created from each Doppler signal<br />
vector. It works by identifying the principal Hankel components that correspond to clutter and<br />
subsequently removing these components to obtain the filtered Doppler signal. On the other<br />
hand, as described in Chapter 3, the Matrix Pencil approach is a parametric spectral estimation<br />
method whose <strong>for</strong>mulation is based on the generalized eigenvalue (GE) problem and low-rank<br />
matrix reduction principles. During its operation, this estimator first creates a rank-reduced<br />
matrix pair from the dominant SVD components of the Hankel data matrix corresponding to a<br />
Doppler signal vector, and then it finds the principal frequencies of the Doppler signal by solving<br />
<strong>for</strong> the GEs of a matrix pencil <strong>for</strong>med from the rank-reduced matrix pair. As can be seen, the<br />
Hankel-SVD filter and the Matrix Pencil approach both begin with the same SVD step in their<br />
theoretical <strong>for</strong>mulation, even though the aim of the two frameworks are somewhat different.<br />
Such similarity in principles makes it possible <strong>for</strong> these two frameworks to be combined together<br />
as a single processing module in the proposed color flow data processor.<br />
4.2.2 Processor Details<br />
Conceptual Overview<br />
To appreciate how the new color flow data processor makes use of the two frameworks<br />
presented earlier in this thesis, we first recall that the Hankel-SVD filter is essentially an adaptive<br />
flow detection approach that involves spectral estimation principles. As described in Section<br />
2.3.2, this approach uses the lag-one autocorrelator to compute the mean Doppler frequency of<br />
each principal Hankel component found from the SVD of the Hankel data matrix. The resulting<br />
spectral estimates are then used to determine the clutter eigen-space dimension: a quantity that is<br />
needed to find the flow signal power as seen in (2-18). Instead of using the lag-one auto-
Chapter 4. Single-Module Approach to <strong>Color</strong> <strong>Flow</strong> Data <strong>Processing</strong> 89<br />
correlator to find the mean frequency of each Hankel component individually, it is possible to<br />
exploit the Hankel matrix structure and apply the Matrix Pencil approach to simultaneously<br />
compute the modal frequencies <strong>for</strong> multiple Hankel components. The advantage of using Matrix<br />
Pencil, as we showed in Chapter 3, is that low-rank matrix reduction can be per<strong>for</strong>med to give<br />
more consistent spectral estimates <strong>for</strong> use during clutter eigen-space analysis. In turn, this<br />
estimator can help the Hankel-SVD filter to more reliably determine the clutter eigen-space<br />
dimension and consequently obtain more accurate flow power estimates. Note that, since the<br />
Matrix Pencil spectral estimates can be used in parallel to find the modal flow velocity (as seen<br />
in (3-18)), it is not necessary to per<strong>for</strong>m separate computations <strong>for</strong> velocity estimation when<br />
using Matrix Pencil as the spectral estimator in the Hankel-SVD filter. In other words, flow<br />
detection and flow estimation can be jointly per<strong>for</strong>med in this combined <strong>for</strong>mulation.<br />
<strong>Based</strong> on the above principles, we developed a new single-module signal processor <strong>for</strong><br />
use in color flow data processing. This new processor, which per<strong>for</strong>ms flow detection and flow<br />
estimation on Doppler ensembles via a Hankel matrix <strong>for</strong>mulation, can be considered as a fusion<br />
between the Hankel-SVD filter’s flow detection concepts and the Matrix Pencil approach’s<br />
multi-modal spectral estimation capabilities. Like the Hankel-SVD filter, the proposed processor<br />
computes the flow signal power via the squared sum of singular values <strong>for</strong> Hankel component<br />
orders higher than the clutter eigen-space dimension. As well, it estimates the modal flow<br />
velocity from the largest frequency in the set of Matrix Pencil spectral estimates obtained during<br />
clutter eigen-space analysis. Note that, in the proposed processor, the clutter eigen-space<br />
dimension is determined by analyzing the spectral spread of the Matrix Pencil frequency<br />
estimates. This analysis approach is similar to the one used to determine the nominal eigenstructure<br />
rank shown in Section 3.3.3. There<strong>for</strong>e, the proposed processor is essentially a<br />
modified rank-adaptive Matrix Pencil estimator that can compute the flow signal power in<br />
addition to the modal flow velocity.<br />
Description of <strong>Processing</strong> Module<br />
A basic input-output diagram of the proposed color flow data processor is illustrated in<br />
Fig. 4-1b. The two major inputs to the processor are the raw Doppler ensemble and the clutter<br />
bandwidth threshold; in return, the processor computes estimates of the flow signal power and<br />
the modal flow velocity. Fig. 4-2 shows the actual sequence of operations per<strong>for</strong>med by the
Chapter 4. Single-Module Approach to <strong>Color</strong> <strong>Flow</strong> Data <strong>Processing</strong> 90<br />
Doppler<br />
<strong>Signal</strong><br />
1. Create Hankel<br />
Matrix A<br />
2. Compute SVD<br />
A = Σσ k<br />
u k<br />
v k<br />
H<br />
3. Initialize Clut<br />
Dim: K c<br />
= 1<br />
4. Form Matrix Pair<br />
With K c<br />
Largest<br />
Hankel Components<br />
[Ã 0<br />
, Ã 1<br />
]<br />
Power Map<br />
Memory<br />
Velocity Map<br />
Memory<br />
8a. Find <strong>Flow</strong> Power<br />
P<br />
Σσ 2<br />
k<br />
K c<br />
ρ =<br />
P(N D –P+1)<br />
8b. Find <strong>Flow</strong> Velocity<br />
v =<br />
2f o cosθ<br />
c o<br />
fD,argmax{|fD,k f D,argmax{|fD,k |}<br />
No<br />
7a. Increment Clut<br />
Dim: K c<br />
=K c<br />
+1<br />
Yes<br />
7. Is Spread<br />
< Thr?<br />
5. Find K c<br />
Modal Freq.<br />
From GEs of à 1<br />
–λà 0<br />
6. Find Spectral Spread<br />
of K c<br />
Modal Freq.<br />
max{f D,k<br />
}–min{f D,k<br />
}<br />
Fig. 4-2. <strong>Flow</strong>chart showing the steps involved in the proposed color flow data processor.<br />
Note that the clutter eigen-space analysis per<strong>for</strong>med in this processor is based on the<br />
adaptive rank selection algorithm used by rank-adaptive Matrix Pencil.<br />
proposed processor to obtain the flow estimates. As can be seen, the proposed processor begins<br />
by computing the principal Hankel components of the Doppler signal via an SVD of the Hankel<br />
data matrix. Note that these initial steps are essentially the same ones as seen in the <strong>for</strong>mulation<br />
of the Hankel-SVD filter and the Matrix Pencil approach. With the SVD components, the<br />
processor then iteratively finds the clutter eigen-space dimension by searching <strong>for</strong> the largest<br />
eigen-structure rank that gives a set of Matrix Pencil frequency estimates with spectral spread<br />
smaller than the specified clutter bandwidth threshold. It is worth pointing out that the clutter<br />
eigen-space dimension found from this analysis is essentially similar to the nominal eigenstructure<br />
rank defined in (3-23). From the clutter eigen-space dimension estimate, the flow<br />
signal power is subsequently computed from the singular values via (2-18). In addition, the<br />
modal flow velocity is found from the set of Matrix Pencil frequency estimates via (3-18). As<br />
seen in this series of steps, a feedback approach is used by the proposed processor to jointly<br />
per<strong>for</strong>m flow detection and flow estimation. With such feedback, more consistent estimates of<br />
the flow signal power and the flow velocity should be obtained as a result.
Chapter 4. Single-Module Approach to <strong>Color</strong> <strong>Flow</strong> Data <strong>Processing</strong> 91<br />
(a) Physical Setting<br />
(b) Theoretical Profile<br />
Fig. 4-3. Imaging scenario <strong>for</strong> the in vitro flow phantom study: (a) an illustration of the<br />
physical setting during data acquisition; (b) a color flow map showing the phantom’s<br />
theoretical flow profile (parabolic flow with a center-line velocity of 70 cm/s).<br />
4.3 In Vitro <strong>Color</strong> <strong>Flow</strong> Imaging Study<br />
4.3.1 Experimental Protocol<br />
Data Acquisition<br />
As an initial per<strong>for</strong>mance assessment, an in vitro flow imaging study was per<strong>for</strong>med to<br />
evaluate the proposed processor’s ability to obtain flow estimates in the presence of motionless<br />
clutter † . As shown in Fig. 4-3a, the imaging view of this experiment is an in-plane slice of a<br />
commercial flow phantom (Gammex-RMI, Middleton, WI, USA; Model 1425A) that has a<br />
5mm-diameter flow tube surrounded by rigid tissue-mimicking material and angled at 50° with<br />
respect to the surface. Note that the bottom-center of the imaging view was aligned with the<br />
flow phantom’s distal tube wall during the data acquisition. <strong>Based</strong> on this physical setting, five<br />
frames of raw color flow data were acquired <strong>for</strong> offline processing using an experimental scanner<br />
(ZONARE Medical Systems, Mountain View, CA, USA) that was equipped with an L10-5<br />
transducer probe. In terms of the acquisition parameters, the color flow dataset was collected by<br />
† The flow phantom setup and the data acquisition portion of this study were per<strong>for</strong>med by Dr. Larry Mo at<br />
ZONARE Medical Systems.
Chapter 4. Single-Module Approach to <strong>Color</strong> <strong>Flow</strong> Data <strong>Processing</strong> 92<br />
Table 4-1. Experimental Parameters in the <strong>Flow</strong> Phantom Study<br />
Parameter<br />
Value<br />
<strong>Flow</strong> Phantom Parameters<br />
Acoustic speed, c o<br />
1540 m/s<br />
Beam-flow angle, θ 50°<br />
Tube diameter<br />
5 mm<br />
<strong>Flow</strong> Profile<br />
Steady, parabolic flow<br />
Center-line flow velocity<br />
70 cm/s<br />
Clutter-to-flow-signal ratio<br />
In the range 20–30 dB<br />
Data Acquisition Parameters<br />
Transmit pulse characteristics<br />
6-cycles, 5MHz sinusoid<br />
Depth sampling rate<br />
5 MHz<br />
Pulse repetition interval, T PRI<br />
175 μs<br />
Doppler ensemble size, N D 14<br />
Image Size Parameters<br />
Lateral field of view –19.2 to +19.2 mm (257 beam lines)<br />
Axial field of view +24.5 to +50.0 mm (166 depth samples)<br />
using a 5-MHz six-cycle sinusoidal pulse along with a 0.175 ms pulse repetition interval and 14<br />
firings per beam line. Also, the online settings of the flow phantom were adjusted to generate a<br />
parabolic flow profile with a center-line velocity of 70 cm/s. A summary of the parameters used<br />
in this in vitro study is provided in Table 4-1.<br />
<strong>Signal</strong> <strong>Processing</strong> Procedure<br />
The first step in the processing of each color flow data frame is to obtain the analytic<br />
Doppler signal of individual sample volumes from the raw ultrasound echoes. For the in vitro<br />
datasets used in this study, such pre-processing was internally carried out in the experimental<br />
scanner’s channel domain memory (Mo et al. 2003). Once the Doppler data ensembles were<br />
derived from the raw echoes, they were passed into the proposed signal processing module to<br />
estimate the flow signal power and flow velocity of each sample volume along all beam lines. In<br />
order to lower the estimation variance, a five-tap median filter and a five-tap mean filter were<br />
applied in sequence to both the flow power and flow velocity estimates. Subsequently, the<br />
smoothed power estimates were color-coded based on their magnitude in a decibel scale, while<br />
the smoothed velocity estimates were color-coded with respect to the unaliased velocity range.<br />
Note that spurious pixels were removed from the power and velocity maps if their corresponding<br />
post-filter signal power was below a given threshold value. In the last step, the non-spurious<br />
pixels were overlaid on top of a B-mode image <strong>for</strong> duplex display.
Chapter 4. Single-Module Approach to <strong>Color</strong> <strong>Flow</strong> Data <strong>Processing</strong> 93<br />
Data Analysis Method<br />
As an analysis benchmark, a theoretical color flow map of the imaging view was first<br />
generated based on the fact that the flow phantom was calibrated to <strong>for</strong>m a parabolic flow profile<br />
with 70 cm/s center-line velocity. This theoretical profile, which is shown in Fig. 4-3b, was then<br />
used to evaluate the proposed processor’s ability to reconstruct the flow dynamics of the<br />
phantom. To facilitate quantitative evaluation, we computed the correlation coefficient between<br />
the theoretical flow profile and each flow velocity map obtained from the proposed processor.<br />
Such per<strong>for</strong>mance measure can be expected to be close to unity if the reconstructed flow maps<br />
were similar to the theoretical flow profile. Besides computation of the correlation coefficient,<br />
we also assessed whether consistent velocity profiles were obtained at all cross-sections of the<br />
flow tube. This analysis was carried out by first locating the flow tube position within the<br />
imaging view (based on the fact that the distal tube wall was at the bottom center and the beamflow<br />
angle was 50°) and subsequently computing the mean and variance of velocity estimates <strong>for</strong><br />
each point across the 5 mm flow tube diameter.<br />
As a comparison <strong>for</strong> the proposed processor, the analysis procedure described above was<br />
repeated <strong>for</strong> two types of color flow data processors that make use of the conventional twomodule<br />
processing approach. The first comparison processor was based on the use of a clutterdownmixing<br />
filter <strong>for</strong> clutter suppression and a lag-one autocorrelator <strong>for</strong> velocity estimation,<br />
and hence it will be hereafter referred to as the “clutter-downmixing processor”. On the other<br />
hand, the second comparison processor was based on the use of a fixed-rank eigen-regression<br />
filter <strong>for</strong> clutter suppression and a lag-one autocorrelator <strong>for</strong> velocity estimation, and it will be<br />
denoted as the “fixed-rank eigen-processor” from hereon. Note that, <strong>for</strong> the eigen-regression<br />
filter of this processor, the Doppler correlation matrix was computed via ensemble averaging of<br />
all the Doppler signal vectors along the same beam line. Such computation approach inherently<br />
presumes that the clutter statistics were stationary over the depth range.<br />
4.3.2 Experimental Results<br />
<strong>Color</strong> <strong>Flow</strong> Images<br />
Fig. 4-4 shows the power maps and velocity maps obtained from the proposed processor<br />
<strong>for</strong> a representative frame in the in vitro dataset. These images were found by using a clutter<br />
bandwidth threshold of 50 Hz <strong>for</strong> the proposed processor, a second-order IIR filter (projection-
Chapter 4. Single-Module Approach to <strong>Color</strong> <strong>Flow</strong> Data <strong>Processing</strong> 94<br />
(a) Proposed Processor<br />
(b) Clutter-Downmixing<br />
Processor<br />
(c) Fixed-Rank<br />
<strong>Eigen</strong>-Processor<br />
Velocity Map<br />
<strong>Flow</strong> Power Map<br />
Corr. Coeff. = 0.978±0.004 Corr. Coeff = 0.972±0.004 Corr. Coeff = 0.979±0.004<br />
Fig. 4-4. <strong>Color</strong> flow images of the flow phantom <strong>for</strong> one of the five frames in the dataset. These<br />
mappings were obtained from: (a) the proposed processor, (b) the clutter-downmixing processor,<br />
and (c) the fixed-rank eigen-processor. The dynamic range of the flow power maps was set to 10<br />
dB, while the unaliased velocity range of the velocity maps was defined to be between –61.6 cm/s<br />
and 75.3 cm/s. Also, the same spurious flow power threshold was used <strong>for</strong> all flow maps.<br />
initialized with 100 Hz nominal cutoff) <strong>for</strong> the clutter-downmixing processor, and a first-order<br />
clutter eigen-space dimension <strong>for</strong> the fixed-rank eigen-processor. From these figures, two<br />
general observations can be made. First, the flow power maps of all three processors have<br />
indicated a presence of flow only within the same spatial region as depicted in the theoretical<br />
profile of Fig. 4-3b. Second, as suggested by their high correlation coefficients (greater than<br />
0.97), the three processors seem to be capable of reconstructing velocity maps that have good<br />
agreement with the theoretical flow profile. These two observations indicate that, <strong>for</strong> imaging<br />
scenarios with no tissue motion, the proposed processor is able to obtain flow estimates that are<br />
comparable to the ones found from conventional two-module processors. Indeed, as will be<br />
subsequently seen in the in vivo studies, the proposed processor can outper<strong>for</strong>m the other two<br />
processors in imaging scenarios when tissue motion is present.<br />
Velocity Profiles<br />
As a more quantitative assessment, Fig. 4-5 shows the cross-sectional velocity profiles<br />
computed from averaging the velocity estimates at corresponding spatial points in the flow tube.
Chapter 4. Single-Module Approach to <strong>Color</strong> <strong>Flow</strong> Data <strong>Processing</strong> 95<br />
(a) Proposed Processor<br />
Proximal Wall<br />
Theory<br />
Estimated<br />
Tube Cross-Section<br />
(5mm)<br />
Distal Wall<br />
70<br />
60<br />
50 40 30 20<br />
<strong>Flow</strong> Velocity [cm/s]<br />
10<br />
0<br />
(b) Clutter-Downmixing Processor<br />
(c) Fixed-Rank <strong>Eigen</strong>-Processor<br />
Proximal Wall<br />
Theory<br />
Estimated<br />
Proximal Wall<br />
Theory<br />
Estimated<br />
Tube Cross-Section<br />
(5mm)<br />
Tube Cross-Section<br />
(5mm)<br />
Distal Wall<br />
Distal Wall<br />
70<br />
60<br />
50 40 30 20<br />
<strong>Flow</strong> Velocity [cm/s]<br />
10<br />
0<br />
70<br />
60<br />
50 40 30 20<br />
<strong>Flow</strong> Velocity [cm/s]<br />
10<br />
0<br />
Fig. 4-5. The flow phantom’s cross-sectional velocity profile as obtained from: (a) the<br />
proposed processor, (b) the clutter-downmixing processor, and (c) the fixed-rank eigenprocessor.<br />
Also shown in these figures are the theoretical profile and the error bars within<br />
one deviation.<br />
As can be seen, the velocity profiles <strong>for</strong> all three processors more or less resemble a parabolic<br />
shape in the central part of the flow tube. Such result is in agreement with the theoretical profile<br />
and helps explain why the flow velocity maps in Fig. 4-4 have a high correlation coefficient.<br />
Nevertheless, it should be noted that the velocity estimates are actually under-biased <strong>for</strong> spatial<br />
points near the tube center and are over-biased <strong>for</strong> spatial points close to the tube wall. This<br />
phenomenon can be explained by recognizing that the spatial points in fact covers a range of<br />
velocities along the parabolic flow gradient (i.e. they are sample volumes); as a result, their
Chapter 4. Single-Module Approach to <strong>Color</strong> <strong>Flow</strong> Data <strong>Processing</strong> 96<br />
modal flow velocity estimates tend to be below maximum near the tube center and above<br />
minimum near the tube wall.<br />
4.4 In Vivo <strong>Color</strong> <strong>Flow</strong> Imaging Studies<br />
4.4.1 Experimental Protocol<br />
Data Acquisition<br />
To facilitate further evaluation of the single-stage processor’s per<strong>for</strong>mance, in vivo color<br />
flow imaging data were acquired from the carotid arteries of a healthy 25-year-old male<br />
volunteer. Each dataset captured the flow dynamics over a full cardiac cycle, and it was<br />
consisted of 15 frames (at 15 Hz frame rate, 1 sec total) of raw ultrasound echoes returned from<br />
different beam lines along the imaging view. During the acquisition of each dataset, the<br />
volunteer was asked to either hold his breath or deeply inhale so that the resulting data frames<br />
would contain non-stationary clutter dominated by either arterial distension or patient movement.<br />
In this in vivo study, all the color flow datasets were obtained using a Philips HDI-5000 scanner<br />
(Bothell, WA, USA) equipped with an L7-4 linear array probe and a special proprietary interface<br />
<strong>for</strong> raw data downloading. The main acquisition parameters of the two in vivo color flow<br />
datasets presented in this section are listed in Table 4-2.<br />
<strong>Signal</strong> <strong>Processing</strong> Procedure<br />
Similar to the in vitro study in Section 4.3, processing of the acquired in vivo data begins<br />
with the conversion of raw echoes into analytic Doppler data samples. Since the HDI-5000<br />
scanner does not per<strong>for</strong>m this conversion internally, an in-phase/quadrature (I/Q) demodulation<br />
routine was used to per<strong>for</strong>m such conversion offline. Data decimation was then carried out after<br />
I/Q demodulation so that the depth sampling rate of each pulse echo was down-converted from<br />
the system’s digital sampling rate to the pulse carrier frequency. Once the Doppler signals were<br />
obtained, they were processed using the proposed processor to compute the flow signal power<br />
and the flow velocity of each sample volume. Prior to display, the flow estimates were passed<br />
through a five-tap median filter and a five-tap mean filter to reduce the estimation variance.<br />
Finally, the smoothed estimates were converted into color pixels based on their respective power<br />
or velocity scale and superimposed onto a B-mode image <strong>for</strong> display. Note that pixels with flow<br />
signal power below a threshold were considered as spurious and were not displayed.
Chapter 4. Single-Module Approach to <strong>Color</strong> <strong>Flow</strong> Data <strong>Processing</strong> 97<br />
Data Analysis Method<br />
In each in vivo color flow imaging study, regions of blood vessel, static tissue, and<br />
moving tissue within each data frame were identified to study the proposed processor’s<br />
sensitivity to blood flow echoes as well as its specificity to stationary and non-stationary Doppler<br />
clutter. For each identified region, the average post-filter signal power was calculated and<br />
compared against each other. Note that, if the proposed processor has good flow detection<br />
capabilities, then its regional power estimate <strong>for</strong> blood vessel sections (where blood flow is<br />
present) should be higher than that <strong>for</strong> tissue sections (where only clutter is present). To study<br />
the proposed processor’s flow detection per<strong>for</strong>mance at various stages in the cardiac cycle, the<br />
regional power analysis was conducted on all frames in the dataset. Also, as a comparative<br />
assessment, the analysis was repeated with the clutter-downmixing processor and the fixed-rank<br />
eigen-processor that were used <strong>for</strong> the in vitro per<strong>for</strong>mance analysis (see Section 4.3.1). It<br />
should be noted that the accuracy of velocity estimates was not analyzed quantitatively in these<br />
in vivo imaging studies since the actual flow dynamics were unknown.<br />
4.4.2 Case Study with Vessel Wall Motion<br />
Overview of Scenario<br />
Table 4-2. Experimental Parameters <strong>for</strong> the In Vivo Imaging Studies<br />
Parameter<br />
Dataset # 1 (while subject<br />
was holding breath)<br />
Data Acquisition Parameters<br />
Dataset # 2 (while subject<br />
was deeply inhaling)<br />
Transmit pulse characteristics<br />
3-cycles, 4MHz sinusoid<br />
Depth sampling rate<br />
4 MHz<br />
Pulse repetition interval, T PRI<br />
333 μs<br />
Doppler ensemble size, N D 10<br />
Number of data frames<br />
15 (at 15 Hz frame rate, 1 sec total)<br />
Acoustic speed, c o<br />
1540 m/s (approximate)<br />
Beam-flow angle, θ<br />
70° (approximate)<br />
Image Size Parameters<br />
Lateral field of view –11.5 to +11.5mm (77 lines) –12.5 to +12.5mm (85 lines)<br />
Axial field of view +5 to +25 mm (115 samples) +5 to +25 mm (115 samples)<br />
We first considered a case where the field of view corresponds to a cross-section of the<br />
carotid arteries beyond bifurcation while the subject was holding his breath (see Fig. 4-6a). In<br />
this case, the tissue motion is primarily caused by arterial wall distension over the cardiac cycle.
Chapter 4. Single-Module Approach to <strong>Color</strong> <strong>Flow</strong> Data <strong>Processing</strong> 98<br />
(a) Physical Setting<br />
(b) B-mode Image<br />
Fig. 4-6. Graphical overview of an in vivo imaging study on the human carotid arteries (past<br />
bifurcation) with tissue motion due to arterial wall distension. In (a), an illustration of the<br />
physical setting is shown. In (b), a B-mode image of the imaging view is shown with<br />
labeled sections of blood vessel (BV), moving tissue (MT), and static tissue (ST). The view<br />
is a cross-sectional slice of the arteries (range of interest marked by green box).<br />
Hence, <strong>for</strong> the regional power analysis, moving tissue sections were selected to be regions<br />
adjacent to the blood vessels (as reflected in the B-mode image in Fig. 4-6b), while a static tissue<br />
section was selected from a region located far away from the blood vessels. For the analysis, a<br />
100 Hz clutter bandwidth threshold was used by the proposed processor since it yielded highpower<br />
flow estimates in the larger blood vessel (the BV1 section in Fig. 4-6b) while giving lowpower<br />
ones in the static tissue section. As well, a second-order filter with 100 Hz nominal cutoff<br />
was used by the IIR-based clutter-downmixing processor to obtain similar power estimates in the<br />
larger blood vessel, while a second-order clutter eigen-space dimension was used by the fixedrank<br />
eigen-processor.<br />
Results: <strong>Color</strong> <strong>Flow</strong> Images<br />
Fig. 4-7a shows the flow power map and the velocity map obtained from the proposed<br />
signal processor design during peak systole where wall motion was the most significant. These<br />
two images should be compared against the ones obtained from the two comparison processors
Chapter 4. Single-Module Approach to <strong>Color</strong> <strong>Flow</strong> Data <strong>Processing</strong> 99<br />
(a) Proposed Processor<br />
(b) Clutter-Downmixing<br />
Processor<br />
(c) Fixed-Rank<br />
<strong>Eigen</strong>-Processor<br />
Velocity Map<br />
<strong>Flow</strong> Power Map<br />
Fig. 4-7. <strong>Color</strong> flow images <strong>for</strong> the in vivo scenario depicted in Fig. 4-6 <strong>for</strong> a frame acquired<br />
during peak systole. The flow power map and velocity map obtained from the proposed<br />
processor are shown in (a), while the ones <strong>for</strong> the two comparison processors are shown in (b)<br />
and (c). The dynamic range of the flow power maps was 15 dB, while the velocity maps had a<br />
range of ±84.4 cm/s. Note that a high color display gain was used <strong>for</strong> these images.<br />
given in Figs. 4-7b and 4-7c. Note that, <strong>for</strong> this series of figures, a high color display gain (i.e. a<br />
low spurious power threshold) was used intentionally to help visualize the processors’ flow<br />
detection per<strong>for</strong>mance. As shown in these images, all three processors have correctly indicated<br />
the presence of flow inside the two blood vessels within the imaging view. Nevertheless, the<br />
flow maps produced by the proposed processor appear to have fewer spurious pixels (often<br />
referred to as flash artifacts) than the ones obtained from the two comparison processors. This<br />
observation suggests that the proposed processor is more capable of distinguishing blood<br />
components from non-stationary Doppler clutter when the processors are optimized to provide<br />
similar flow power estimates in the larger blood vessel. Another point worth noting is that,
Chapter 4. Single-Module Approach to <strong>Color</strong> <strong>Flow</strong> Data <strong>Processing</strong> 100<br />
amongst the three processors considered in this study, the fixed-rank eigen-processor actually<br />
produced flow maps with the most spurious pixels. The mediocre per<strong>for</strong>mance of the fixed-rank<br />
eigen-processor is likely due to two factors: 1) the use of a fixed clutter eigen-space dimension<br />
may not be effective since clutter statistics generally vary over time; 2) estimation of the Doppler<br />
correlation matrix via multi-depth ensemble averaging may be impractical given the spatial<br />
varying nature of clutter characteristics over a depth range of few centimeters. <strong>Based</strong> on these<br />
problems, it seems that modifications to the <strong>for</strong>mulation of the fixed-rank eigen-processor are<br />
needed in order <strong>for</strong> the processor to be effective in in vivo imaging scenarios where tissue motion<br />
is generally present † .<br />
Results: Regional Power Comparison<br />
To provide further insights on the efficacy of the color flow data processors, Figs. 4-8a<br />
through 4-8c show the average post-filter Doppler power <strong>for</strong> different regions in the imaging<br />
view as a function of the frame number (with peak systole happening at frame # 6). In addition,<br />
Fig. 4-8d shows a cross-processor comparison on the power ratios between the larger vessel<br />
(BV1) and its adjacent moving tissue region (MT1). From these plots, it can be seen that during<br />
peak systole (around frame # 6), all three data processors considered in this study can more or less<br />
obtain flow power estimates that are distinguishable from the post-filter tissue power estimates;<br />
during cardiac diastole (i.e. other data frames besides frame # 6), a similar result can be seen as<br />
well. Nonetheless, as reflected in the figures, the proposed processor appears to have a better<br />
separation between the power estimates of blood and tissue regions during cardiac systole where<br />
substantial wall motion was present. This observation helps explain why the color flow images<br />
obtained from the proposed processor have fewer spurious pixels than the ones found from the<br />
other two processors.<br />
4.4.3 Case Study with Translational Tissue Motion<br />
Overview of Scenario<br />
In the second case study, we considered a scenario where both tissue and blood vessel<br />
were progressively pushed towards the transducer throughout the frame acquisition period. This<br />
translational tissue motion was caused by the deep inhalation that the subject undertook while<br />
† To our knowledge, an investigation of the problems associated with the eigen-regression filter is currently being<br />
pursued as a doctoral study in the Norwegian research group (e.g. see Lovstakken et al. 2006).
Chapter 4. Single-Module Approach to <strong>Color</strong> <strong>Flow</strong> Data <strong>Processing</strong> 101<br />
<strong>Signal</strong> Power [dB]<br />
(a) Regional Power (Proposed)<br />
BV1 BV2 MT1 MT2 ST1<br />
Relative Time<br />
60<br />
0 0.2 0.4 0.6<br />
55<br />
0.8 1<br />
50<br />
45<br />
40<br />
35<br />
30<br />
25<br />
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15<br />
Frame Num<br />
(c) Regional Power (Fixed-Rank <strong>Eigen</strong>)<br />
BV1 BV2 MT1 MT2 ST1<br />
(b) Regional Power (Clutter-Downmixing)<br />
BV1 BV2 MT1 MT2 ST1<br />
<strong>Signal</strong> Power [dB]<br />
Relative Time<br />
60<br />
0 0.2 0.4 0.6<br />
55<br />
0.8 1<br />
50<br />
45<br />
40<br />
35<br />
30<br />
25<br />
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15<br />
Frame Num<br />
(d) Power Ratio: BV1/MT1<br />
Proposed Downmix Fixed-<strong>Eigen</strong><br />
<strong>Signal</strong> Power [dB]<br />
Relative Time<br />
60<br />
0 0.2 0.4 0.6<br />
55<br />
0.8 1<br />
50<br />
45<br />
40<br />
35<br />
30<br />
25<br />
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15<br />
Frame Num<br />
BV1/MT1 Ratio [dB]<br />
0 0.2<br />
Relative Time<br />
0.4 0.6 0.8 1<br />
20<br />
15<br />
10<br />
5<br />
0<br />
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15<br />
Frame Num<br />
Fig. 4-8. Regional post-filter power estimates <strong>for</strong> different flow image frames acquired <strong>for</strong><br />
the in vivo scenario considered in Fig. 4-6. Results are shown <strong>for</strong>: (a) the proposed<br />
processor, (b) the clutter-downmixing processor, and (c) the fixed-rank eigen-processor. In<br />
(d), the power ratio between the BV1 and MT1 sections are shown <strong>for</strong> the three processors.<br />
Note that peak systole occurs at Frame # 6 in this sequence (marked by gray vertical line).<br />
color flow data was acquired. It essentially masked out the arterial wall distensions and gave rise<br />
to Doppler clutter that was mostly non-stationary throughout the cardiac cycle.<br />
As illustrated in Fig. 4-9a, the imaging view of this second case study is an in-plane slice<br />
of the subject’s common carotid artery. It is important to note that, during the data acquisition,<br />
most of the imaging view was affected by the translational tissue motion. Hence, when<br />
analyzing the post-filter power estimates, the two adjacent regions above and below the blood<br />
vessel were labeled as moving tissue regions (see Fig. 4-9b), while only a small region near the<br />
top edge of the color box was labeled as static tissue.
Chapter 4. Single-Module Approach to <strong>Color</strong> <strong>Flow</strong> Data <strong>Processing</strong> 102<br />
(a) Physical Setting<br />
(b) B-mode Image<br />
Fig. 4-9. Illustration of a flow study on a human common carotid artery in the presence of<br />
translational tissue motion. The physical setting of this imaging scenario is shown in (a),<br />
and a labeled B-mode image of the imaging view is shown in (b). Note that the view is an<br />
in-plane slice of a subject’s common carotid artery.<br />
In processing this dataset, we found that a clutter bandwidth threshold of 150 Hz was<br />
needed by the proposed processor to discriminate between blood and clutter components while<br />
maintaining reasonable flow power estimates within the blood vessel. Also, <strong>for</strong> the comparison<br />
processors, it was found that the IIR-based clutter-downmixing processor needed to use a thirdorder<br />
filter with 150 Hz nominal cutoff, whereas the fixed-rank eigen-processor required a thirdorder<br />
clutter eigen-space dimension. Note that these threshold values and filter stopband sizes<br />
are larger than the one used in the previous case study, but such increase can be expected since<br />
non-stationary Doppler clutter is often more wideband in nature.<br />
Results: <strong>Color</strong> <strong>Flow</strong> Images<br />
To examine the worst-case scenario <strong>for</strong> this case study, we focused on a data frame<br />
acquired during cardiac diastole where blood and clutter components were less separated in<br />
Doppler frequency due to the slower blood flow. In Fig. 4-10, the color flow images<br />
corresponding to such data frame are shown <strong>for</strong> the proposed processor and the two comparison<br />
processors. Like the previous case study, a high color display gain was purposely used <strong>for</strong> these
Chapter 4. Single-Module Approach to <strong>Color</strong> <strong>Flow</strong> Data <strong>Processing</strong> 103<br />
(a) Proposed Processor<br />
(b) Clutter-Downmixing<br />
Processor<br />
(c) Fixed-Rank<br />
<strong>Eigen</strong>-Processor<br />
Velocity Map<br />
<strong>Flow</strong> Power Map<br />
Fig. 4-10. <strong>Color</strong> flow images <strong>for</strong> the in vivo scenario shown in Fig. 4-8 <strong>for</strong> a data frame<br />
acquired during cardiac diastole. In (a), the flow power map and velocity map <strong>for</strong> the<br />
proposed processor are shown. In (b) and (c), the corresponding flow maps <strong>for</strong> the two<br />
comparison processors are shown respectively. The dynamic range and unaliased velocity<br />
range are the same as the ones used in Fig. 4-7. Also, a high color display gain was used <strong>for</strong><br />
these images.<br />
images to help visualize the flow detection per<strong>for</strong>mance of the three processors. As can be seen,<br />
the quality of these color flow maps is generally degraded by the presence of many spurious<br />
pixels. This result indicates that all the color flow data processors had suffered a drop in the<br />
flow detection per<strong>for</strong>mance to some extent. Nevertheless, from the images, it can be seen that<br />
the proposed processor have produced flow maps with fewer spurious pixels than the ones<br />
obtained using the two comparison processors. Indeed, amongst the three processors, the<br />
proposed processor seems to be the only one that can give flow maps with color pixels located<br />
mainly inside the blood vessel. This observation suggests that the proposed processor has better
Chapter 4. Single-Module Approach to <strong>Color</strong> <strong>Flow</strong> Data <strong>Processing</strong> 104<br />
<strong>Signal</strong> Power [dB]<br />
(a) Regional Power (Proposed)<br />
BV1 MT1 MT2 ST1<br />
Relative Time<br />
60<br />
0 0.2 0.4 0.6<br />
55<br />
0.8 1<br />
50<br />
45<br />
40<br />
35<br />
30<br />
25<br />
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15<br />
Frame Num<br />
(c) Regional Power (Fixed-Rank <strong>Eigen</strong>)<br />
BV1 MT1 MT2 ST1<br />
(b) Regional Power (Clutter-Downmixing)<br />
BV1 MT1 MT2 ST1<br />
<strong>Signal</strong> Power [dB]<br />
Relative Time<br />
60<br />
0 0.2 0.4 0.6<br />
55<br />
0.8 1<br />
50<br />
45<br />
40<br />
35<br />
30<br />
25<br />
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15<br />
Frame Num<br />
(d) Power Ratio: BV1/MT2<br />
Proposed Downmix Fixed-<strong>Eigen</strong><br />
<strong>Signal</strong> Power [dB]<br />
Relative Time<br />
60<br />
0 0.2 0.4 0.6<br />
55<br />
0.8 1<br />
50<br />
45<br />
40<br />
35<br />
30<br />
25<br />
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15<br />
Frame Num<br />
flow visualization per<strong>for</strong>mance than the two comparison processors in the worst-case scenario of<br />
the current in vivo study.<br />
Results: Regional Power Comparison<br />
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15<br />
Frame Num<br />
As an analysis of the flow detection per<strong>for</strong>mance at various cardiac cycle phases in this<br />
case study, Figs. 4-11a to 4-11c show the regional post-filter power estimates of the proposed<br />
processor and the two comparison processors <strong>for</strong> different data frames (with systole occurring at<br />
the start and end of sequence). Also, a cross-processor comparison is provided in Fig. 4-11d <strong>for</strong><br />
BV1/MT2 Ratio [dB]<br />
20<br />
15<br />
10<br />
5<br />
0<br />
Relative Time<br />
0 0.2 0.4 0.6 0.8 1<br />
Fig. 4-11. Regional post-filter power estimates <strong>for</strong> all the image frames in the in vivo<br />
scenario considered in Fig. 4-8. Results <strong>for</strong> the proposed processor, the clutter-downmixing<br />
processor, and the fixed-rank eigen-processor are respectively shown in (a), (b), and (c).<br />
For reference, the power ratios between the BV1 and MT2 sections are shown in (d) <strong>for</strong> the<br />
three processors. Note that peak systole occurs at Frames # 2 and # 14 of this sequence<br />
(marked by the gray vertical lines).
Chapter 4. Single-Module Approach to <strong>Color</strong> <strong>Flow</strong> Data <strong>Processing</strong> 105<br />
the power ratio between the blood vessel section (BV1) and the more significant moving tissue<br />
section (MT2). These figures indicate that during the cardiac systolic phase (near frames # 2 and<br />
# 14), all three processors are generally able to distinguish blood signals from Doppler clutter. In<br />
contrast, during cardiac diastole (between frames # 2 and # 14), the proposed processor is more<br />
capable of discriminating between blood and tissue regions, as both the clutter-downmixing<br />
processor and the fixed-rank eigen-processor seem to have difficulties in preserving blood<br />
signals while suppressing Doppler clutter. This observation is consistent with the insights drawn<br />
from the color flow images seen in Fig. 4-10.<br />
4.5 Concluding Remarks<br />
<strong>Color</strong> flow signal processing is conventionally per<strong>for</strong>med using a two-module approach<br />
where flow detection is first carried out prior to flow estimation. The primary shortcoming of<br />
this two-module processing approach, however, is that the flow estimation accuracy inevitably<br />
depends on the efficacy of the flow detection module. To avoid problems with the two-module<br />
processing approach, this chapter has described a new single-module processor design that can<br />
per<strong>for</strong>m flow detection and flow estimation in parallel. The proposed processor is based on the<br />
combination of detection and estimation principles drawn from the Hankel-SVD filter and the<br />
Matrix Pencil estimator, which are the two frameworks described in the previous chapters.<br />
In our steady-flow phantom imaging study, the proposed processor has demonstrated its<br />
ability in reconstructing velocity maps that more or less follow the theoretical profile. Also, <strong>for</strong><br />
this in vitro scenario where the clutter is stationary, the proposed processor has shown that its<br />
flow estimation accuracy and flow detection per<strong>for</strong>mance are similar to two kinds of<br />
conventional two-module processors. On the other hand, when this new processor is applied to<br />
in vivo imaging scenarios with non-stationary clutter, it appears to have better flow detection<br />
capabilities than the conventional processors. In particular, fewer spurious pixels are seen in the<br />
flow maps obtained from the new processor, and more favorable regional power estimates are<br />
observed in the corresponding blood and tissue sections of the color flow images. From these<br />
findings, it seems that the single-module processor presented in this chapter has higher potential<br />
than conventional ones in distinguishing blood flow signals from Doppler clutter.
CHAPTER 5<br />
Thesis Summary and Future Directions<br />
5.1 Chapter Overview<br />
The purpose of this short concluding chapter is to summarize the key features of the new<br />
color flow signal processing methods presented in the previous three chapters. In particular, it is<br />
our intent to review the theoretical advantages of the proposed methods over existing ones. In<br />
addition to the summary, we will also provide some suggestions on the potential future research<br />
directions that can be pursued along this line of work. As such, the rest of this chapter has been<br />
organized as follows:<br />
• Section 5.2 reviews the significance of this research work on color flow signal processing<br />
and lists the major research contributions;<br />
• Section 5.3 outlines some prospective research topics on three different aspects of color flow<br />
signal processing and suggests potential ways to approach these topics;<br />
• Section 5.4 gives some concluding remarks on the potential impact of this research.<br />
5.2 Summary of Thesis Study<br />
5.2.1 Overall Scope<br />
This thesis study has been devoted to the design of new eigen-based signal processing<br />
methods <strong>for</strong> use in color flow imaging. In particular, it has been our intent to develop a new<br />
eigen-based clutter filter and a new parametric flow estimator to effectively account <strong>for</strong> three<br />
main problems in color flow signal processing: 1) the lack of abundant Doppler samples<br />
available <strong>for</strong> processing, 2) the possible presence of wideband Doppler clutter due to tissue<br />
motion, and 3) the potential flow signal distortions that may arise during clutter suppression. In<br />
approaching this task, we made use of the eigen-space principles related to two matrix <strong>for</strong>ms<br />
known as the Hankel matrix and the matrix pencil. To our knowledge, such <strong>for</strong>mulation has not<br />
been introduced be<strong>for</strong>e in the research field of color flow signal processing. Indeed, it is our<br />
- 106 -
Chapter 5. Thesis Summary and Future Directions 107<br />
understanding that none of the existing clutter filters and flow estimators involves the Hankel<br />
matrix or the matrix pencil as part of their <strong>for</strong>mulation, even though a few types of eigen-based<br />
signal processing methods have been proposed previously.<br />
Common to the signal processing methods proposed in this thesis is their adaptability to<br />
the Doppler signal contents. Such adaptability is achieved via a singular value decomposition<br />
(SVD) analysis of the Hankel matrix created from each Doppler signal vector. It is worth noting<br />
that, because of their <strong>for</strong>mulation through the Hankel data matrix, all of the proposed methods<br />
are intended to work with each Doppler ensemble separately. This <strong>for</strong>mulation should be<br />
distinguished from existing eigen-based color flow data processing methods that require Doppler<br />
ensembles from multiple sample volumes to estimate the correlation matrix statistics. In fact, the<br />
proposed eigen-based methods should be more useful than the existing ones because multiple<br />
sample volumes with similar Doppler signal characteristics may not always be available<br />
(especially when spatially-varying tissue motion is present).<br />
In our study, the potential of the proposed color flow signal processing methods were<br />
examined using both simulation and experimental means. For the simulations, we developed an<br />
in-house Doppler signal synthesis model to generate Doppler signals with non-stationary clutter<br />
characteristics and transit-time-broadened blood flow echoes. For the experiments, we acquired<br />
color flow imaging datasets from a steady-flow phantom and the human carotid arteries. From<br />
these analyses, it was generally observed that the proposed color flow signal processing methods<br />
have better capabilities than existing ones in discriminating blood flow signals from Doppler<br />
clutter and in obtaining flow estimates that are less biased. Hence, the proposed methods can<br />
potentially improve the flow visualization per<strong>for</strong>mance of ultrasound color flow images.<br />
5.2.2 Specific Developments<br />
Hankel-SVD Filter<br />
The first signal processing method that we developed in this thesis is a new eigen-based<br />
clutter suppression method called the Hankel-SVD filter. This new clutter filter design, as<br />
described in Chapter 2, is adaptive to the Doppler signal contents, and it is derived using the<br />
notion of principal Hankel component analysis. Similar to existing eigen-regression filters<br />
(Bjaerum et al. 2002b), the Hankel-SVD filter attempts to decompose the Doppler signal as a<br />
series of adaptable and orthogonal components. However, instead of estimating the components
Chapter 5. Thesis Summary and Future Directions 108<br />
through an eigen-decomposition of the Doppler correlation matrix <strong>for</strong>med from multiple signal<br />
snapshots, the new filter computes these components through the SVD of a Hankel data matrix<br />
created from each Doppler signal. To achieve clutter suppression, the filter then removes the<br />
principal Hankel components that correspond to tissue clutter by analyzing their mean Doppler<br />
frequency. Note that, <strong>for</strong> this filter design, we have shown that the filtered Doppler power (i.e.<br />
the flow signal power when blood flow is present) can actually be estimated directly from the<br />
squared sum of singular values <strong>for</strong> the non-clutter Hankel components. There<strong>for</strong>e, it is not<br />
necessary to directly reconstruct the filtered Doppler signal vector.<br />
To study the efficacy of the Hankel-SVD filter, simulations were carried out via the use<br />
of an in-house Doppler signal synthesis model (to be described in the last part of this section). In<br />
these simulations, multiple sets of Doppler data were synthesized <strong>for</strong> two scenarios: 1) a case<br />
with arterial flow (>5cm/s), slowly moving tissue (
Chapter 5. Thesis Summary and Future Directions 109<br />
Doppler signal. Hence, by including the clutter harmonics as part of the signal eigen-structure,<br />
this framework can be used to extract the mode flow velocity in the presence of clutter. <strong>Based</strong><br />
on this notion, we have developed several <strong>for</strong>ms of Matrix-Pencil-based flow estimators that can<br />
potentially be useful to color flow data processing. These include both fixed-rank flow<br />
estimators (rank-one/rank-two Matrix Pencil) as well as adaptive-rank flow estimators (rankadaptive<br />
Matrix Pencil). Note that, <strong>for</strong> the adaptive-rank flow estimator, the nominal eigenstructure<br />
rank was found using a new rank-selection algorithm that is based on the spectral<br />
spread of principal frequency estimates.<br />
To assess the per<strong>for</strong>mance of the Matrix Pencil flow estimators, we used our in-house<br />
Doppler signal synthesis model to generate datasets <strong>for</strong> two flow scenarios: a case with<br />
narrowband Doppler clutter (no tissue motion) and another with wideband Doppler clutter (up to<br />
1 cm/s tissue velocity). Note that the flow velocity in the simulations ranged between 0-15 cm/s,<br />
and the ensemble period was set at 10 ms. <strong>Based</strong> on these synthesized datasets, the estimation<br />
bias and the root-mean-squared error of the Matrix Pencil flow estimators were computed as a<br />
function of blood-signal-to-noise ratio and blood velocity. Results have indicated that the Matrix<br />
Pencil flow estimators are generally less biased than most of the existing frequency-based flow<br />
estimators such as the lag-one autocorrelator. As such, the Matrix Pencil method seems capable<br />
of improving the accuracy of flow estimates.<br />
Single-Module <strong>Color</strong> <strong>Flow</strong> Data Processor<br />
<strong>Based</strong> on the Hankel-SVD filter and the Matrix Pencil method, we have developed a new<br />
color flow signal processor to per<strong>for</strong>m flow detection and flow estimation in parallel within the<br />
same processing module. As described in Chapter 4, this new processor works by combining the<br />
frequency-based flow detection concepts from the Hankel-SVD filter and the multi-modal<br />
spectral estimation capabilities in the Matrix Pencil method. Specifically, it computes the flow<br />
signal power in the same way as done in the Hankel-SVD filter (i.e. by calculating the squared<br />
sum of the non-clutter singular values), and it estimates the modal flow velocity in the same way<br />
as carried out in the Matrix Pencil method (i.e. by finding the largest frequency in the set of<br />
principal spectral estimates). Note that, <strong>for</strong> this single-module estimator, the clutter eigen-space<br />
dimension was determined using a frequency-based feedback approach that iteratively searches<br />
<strong>for</strong> the largest eigen-structure rank with a spectral spread smaller than the specified clutter
Chapter 5. Thesis Summary and Future Directions 110<br />
bandwidth threshold (just like the rank-adaptive Matrix Pencil). Such a feedback approach is<br />
intended to provide more consistent estimates of the flow signal power and the flow velocity.<br />
To examine its per<strong>for</strong>mance, the single-module processor was applied to color flow data<br />
acquired from in vitro and in vivo imaging scenarios. First, this processor was used to compute<br />
flow estimates from an in vitro dataset corresponding to a steady-flow phantom (parabolic flow,<br />
70 cm/s center-line velocity) with tissue-mimicking material surrounding the flow tube. For this<br />
study, it was found that the single-module processor can obtain flow power maps that resemble<br />
the flow tube’s shape and can reconstruct velocity maps that have high correlation with the<br />
theoretical flow profile. Aside from the in vitro study, the single-module processor was also<br />
applied to two in vivo imaging scenarios where the field of view corresponds to the human<br />
carotid arteries when different degrees of tissue motion were present. It was found that this<br />
processor can produce color flow images with fewer spurious pixels than the ones obtained from<br />
conventional two-module processors. These results suggest that the single-module processor is<br />
generally more effective than conventional ones in identifying blood flow in the presence of<br />
clutter (especially non-stationary ones).<br />
Doppler <strong>Signal</strong> Synthesis Model<br />
For the purpose of analyzing the theoretical per<strong>for</strong>mance of the Hankel-SVD filter and<br />
the Matrix Pencil estimator, a Doppler signal synthesis model was developed as part of the<br />
background work in this thesis. As described in Section 2.4, our synthesis model separately<br />
generates clutter and blood components of the Doppler signal based on two different signal<br />
models. In particular, the Doppler clutter is synthesized via a phase-modulated signal model,<br />
while blood echoes are generated using an amplitude-modulated signal model. The use of two<br />
different models to simulate Doppler signal is physically justified by recognizing that tissue and<br />
blood scatterers follow different movement mechanisms: tissue tends to move in a quasi-cyclic<br />
pattern while blood scatterers simply traverse through the sample volume. Note that, in the<br />
literature, such a hybrid approach <strong>for</strong> synthesizing Doppler data has not been considered be<strong>for</strong>e.<br />
Indeed, existing Doppler signal synthesis approaches often treat both clutter and blood echoes as<br />
amplitude-modulated signals (e.g. see Bjaerum et al. 2002a), despite the fact that amplitude<br />
modulation does not seem to be an appropriate way of modeling tissue clutter’s movement<br />
mechanism.
Chapter 5. Thesis Summary and Future Directions 111<br />
5.2.3 List of Primary Research Contributions<br />
The original contributions of this thesis are mainly in the <strong>for</strong>m of theoretical development<br />
along with experimental assessment. In particular, the following is a list of the primary research<br />
contributions made in this body of work:<br />
(a) Development of a new class of eigen-based color flow signal processing methods that are<br />
intended to work with the Doppler signal of each sample volume separately;<br />
(b) Design of a new clutter filter (Hankel-SVD filter) that is adaptive to the Doppler signal<br />
contents via the use of an existing framework called principal Hankel component analysis;<br />
(c) Development of a new eigen-based velocity estimator (Matrix Pencil) based on the eigenspace<br />
principles of an existing array processing framework called the Matrix Pencil;<br />
(d) Creation and implementation of a new Doppler signal synthesis model that can simulate nonstationary<br />
Doppler clutter and transit-time-broadened blood flow echoes;<br />
(e) Analysis of the Hankel-SVD filter’s flow detection per<strong>for</strong>mance and the Matrix Pencil<br />
method’s flow estimation accuracy via the use of our simulation model mentioned in (d);<br />
(f) Proposal of a novel single-module color flow data processor that per<strong>for</strong>ms flow detection and<br />
flow estimation in parallel via the combination of principles from the Hankel-SVD filter and<br />
the Matrix Pencil estimator;<br />
(g) Offline demonstration of the single-module processor’s improved capability over existing<br />
ones to detect blood flow in in vivo imaging scenarios with significant tissue motion (i.e.<br />
non-stationary Doppler clutter).<br />
Note that point (a) can be considered as the overall contribution of this thesis, while the other six<br />
points can be considered as specific contributions.<br />
5.3 Future Research Directions<br />
5.3.1 Developmental Aspects<br />
Real-Time Computation Issues<br />
To extend beyond the findings presented in this thesis, there are quite a few research<br />
topics that are worth exploring in the future. One of these prospective topics is on addressing the<br />
computational aspect of eigen-based signal processing strategies. In particular, it is wellrecognized<br />
that the primary challenge <strong>for</strong> implementing eigen-based signal processors in real-
Chapter 5. Thesis Summary and Future Directions 112<br />
time is the high computational complexity involved with the SVD or eigen-decomposition step<br />
(which require on the order of P 3 flops). Hence, a more efficient computation algorithm should<br />
be developed to help reduce the overall complexity of eigen-based methods, thereby making<br />
them more feasible <strong>for</strong> real-time implementation. For the case of the Hankel-SVD filter and the<br />
Matrix Pencil estimator, a potential approach <strong>for</strong> designing such algorithm is to exploit the<br />
persymmetric structure of the Hankel matrix during SVD computation. This design approach<br />
has been previously explored in a few theoretical studies (e.g. see Bedeau et al. 2004), and thus it<br />
will be worthwhile to examine its applicability to color flow signal processing (possibly in<br />
collaboration with experts in the matrix computation field).<br />
Adaptive Thresholding<br />
Another topic that can be pursued in the future is on developing an algorithm to<br />
adaptively select the clutter bandwidth threshold used by the single-module processor proposed<br />
in this thesis. The motivations <strong>for</strong> designing such an algorithm can be perceived by first<br />
recalling that the clutter bandwidth threshold is the main parameter influencing the flow<br />
detection/estimation per<strong>for</strong>mance of the single-module processor. In our study, this threshold<br />
was designated as a user-defined parameter in which the user can adjust between zero and the<br />
Doppler aliasing frequency, and the same threshold value was used to process the Doppler data<br />
in all the sample volumes within the imaging view. As such, the availability of an adaptivethresholding<br />
algorithm should further improve the proposed processor’s per<strong>for</strong>mance in<br />
scenarios where the Doppler clutter bandwidth varies significantly over both space and time.<br />
Note that, in developing this algorithm, one of the features in which the adaptation can be based<br />
upon is the Doppler clutter’s center frequency. Specifically, it can be anticipated that the clutter<br />
bandwidth is larger when the clutter spectrum deviates substantially from zero frequency since<br />
tissue motion is more likely to be significant in this case. Such property was exploited in a<br />
recent investigation by Yoo et al. (2003) on adaptive clutter filter selection.<br />
5.3.2 Experimental Aspects<br />
<strong>Flow</strong> Imaging Phantom Design<br />
An experimental research direction that can be pursued in the future is on designing a<br />
flow imaging phantom that has an elastic flow tube padded with a layer of tissue mimicking<br />
material. The aim of developing such a phantom is to facilitate rational modeling of the Doppler
Chapter 5. Thesis Summary and Future Directions 113<br />
echoes originating from a pulsating artery surrounded by body tissues, thereby allowing us to<br />
acquire experimental color flow data with more realistic clutter characteristics and in turn study<br />
the per<strong>for</strong>mance of color flow signal processors as a function of tissue elasticity and flow<br />
parameters. It is worth noting that, in the ultrasound research field, few have considered the<br />
design of flow imaging phantoms with elastic tissues and flow tubes. Indeed, like the Gammex-<br />
RMI phantom used in our study, most existing flow imaging phantoms mainly comprise a rigid<br />
flow tube that is positioned below a layer of tissue-like gel. There<strong>for</strong>e, development of an elastic<br />
flow imaging phantom can be a substantial contribution to flow imaging research and system<br />
testing. Note that Kargel et al. (2003) have recently reported a dual-tube flow imaging phantom<br />
design in which one of the tubes is made of elastic material. This design appears to be an<br />
encouraging first step towards the development of an elastic flow imaging phantom.<br />
Doppler Clutter Analysis<br />
In parallel to the design of a flow imaging phantom, it may be worth conducting a<br />
complementary study on the Doppler clutter characteristics <strong>for</strong> various in vivo scenarios. In<br />
particular, it should be useful to analyze how tissue motion can affect the Doppler clutter <strong>for</strong><br />
different sample volumes within the imaging view and at different parts of a cardiac cycle. One<br />
way of per<strong>for</strong>ming such an analysis is to study the statistical distribution of the main Doppler<br />
clutter features like energy, bandwidth, and center frequency. Correspondingly, new statistical<br />
models and mathematical approximations may be developed to predict the space-time variations<br />
of Doppler clutter in color flow data, thereby allowing us to more appropriately define the clutter<br />
stopband of color flow signal processors and design more suitable adaptive-thresholding<br />
algorithms. It should be noted that such a space-time statistical relationship <strong>for</strong> Doppler clutter<br />
has not been reported in the literature, and perhaps the closest of such kind is a phase-modulated<br />
clutter model that was developed to describe the clutter behavior in spectral Doppler studies<br />
(Heimdal and Torp 1997). Hence, this research direction can potentially lead to significant<br />
contributions to the current understanding of Doppler clutter.<br />
5.3.3 Theoretical Aspects<br />
<strong>Processing</strong> of Non-Uni<strong>for</strong>mly Sampled <strong>Color</strong> <strong>Flow</strong> Data<br />
Aside from experimental research topics, some theoretical topics in color flow signal<br />
processing are also worth exploring upon in future studies. One of these topics is related to the
Chapter 5. Thesis Summary and Future Directions 114<br />
processing of non-uni<strong>for</strong>mly sampled color flow data. As recently demonstrated in a preliminary<br />
study (Wilkening et al. 2002), non-uni<strong>for</strong>m color flow data sampling has the advantage of<br />
virtually extending the aliasing limit without losing velocity resolution, but its resulting flow<br />
estimates tend to be less accurate than the ones derived from uni<strong>for</strong>mly sampled data. It is thus<br />
worthwhile to develop a more robust flow estimator in order <strong>for</strong> this non-uni<strong>for</strong>m sampling<br />
technique to obtain accurate flow estimates. One potential way of designing such an estimator is<br />
to modify the Matrix Pencil approach considered in this thesis so that the framework can be used<br />
to process non-uni<strong>for</strong>mly sampled data. Note that the estimator design may also be based on<br />
solutions adopted from the array processing literature, since it is well-known that spectral<br />
estimation with non-uni<strong>for</strong>mly sampled data is essentially the same as direction-of-arrival<br />
estimation using a non-uni<strong>for</strong>mly spaced antenna array.<br />
Contrast-Enhanced <strong>Color</strong> <strong>Flow</strong> Data <strong>Processing</strong><br />
The use of parametric methods to account <strong>for</strong> nonlinear harmonics in contrast-enhanced<br />
color flow data is another theoretical topic that can be studied in the future. As described by<br />
Bruce et al. (2004), the presence of nonlinear harmonics in the pulse echoes acquired using<br />
microbubble contrast agents (in conjunction with a pulse-inversion firing scheme) is the main<br />
factor that complicates the processing of contrast-enhanced color flow data. From a twodimensional<br />
(2D) frequency perspective, these nonlinear harmonics give rise to multiple<br />
components on the fast-time/slow-time frequency space even <strong>for</strong> single-velocity flow profiles,<br />
and thus multi-stage filtering strategies are typically needed to process contrast-enhanced color<br />
flow data. As a more direct approach to account <strong>for</strong> the nonlinear harmonics, it may be possible<br />
to use 2D parametric estimation methods to simultaneously compute all the principal modes in<br />
the fast-time/slow-time frequency space. For instance, a 2D parametric estimator that can<br />
potentially be useful is the 2D <strong>for</strong>mulation of Matrix Pencil (Hua 1992). Hopefully, the use of<br />
2D parametric estimators can simplify the conventional multi-stage approach to the processing of<br />
contrast-enhanced color flow data.<br />
Other Single-Snapshot-<strong>Based</strong> <strong>Flow</strong> Estimators<br />
As pointed out earlier, this thesis is the first attempt in the ultrasound research<br />
community to per<strong>for</strong>m eigen-based, parametric flow estimation without involving the use of<br />
Doppler ensembles from multiple sample volumes. In approaching this task, we have made use
Chapter 5. Thesis Summary and Future Directions 115<br />
of the Matrix Pencil framework that is well-recognized in the array processing field and have<br />
developed three <strong>for</strong>ms of flow estimators using such method. It is worth noting that the Matrix<br />
Pencil is not the only kind of single-snapshot-based parametric estimation method available,<br />
though this approach was chosen <strong>for</strong> use in our study because of its computational efficiency as<br />
compared to other parametric estimators. In the future, it will be worthwhile to study the<br />
applicability of other single-snapshot parametric estimation methods to color flow signal<br />
processing. One particular kind of these methods worth devoting attention to is the deterministic<br />
maximum likelihood method, which theoretically has a more precise estimation per<strong>for</strong>mance<br />
than the Matrix Pencil approach as demonstrated in array processing field (van der Veen et al.<br />
1993). It would be interesting to compare the per<strong>for</strong>mance of various parametric estimators in<br />
the context of color flow signal processing.<br />
<strong>Flow</strong> Analysis Using Image <strong>Processing</strong> Tools<br />
Instead of approaching color flow data processing from a one-dimensional signal analysis<br />
perspective, it may be possible to carry out the same task by making use of image processing<br />
tools that are well-established in the computer vision field. Indeed, as reviewed by Hein and<br />
O’Brien (1993), such an image-processing-based flow analysis approach has been attempted in<br />
some heuristic studies where simple motion-tracking methods like block matching and optical<br />
flow analysis have been used to examine the inter-pulse movement of blood scatterers. Although<br />
these early studies did not achieve favorable per<strong>for</strong>mance results under in vivo flow scenarios<br />
(because of problems caused by tissue motion), they have still demonstrated that image<br />
processing tools may offer an alternative way of computing flow estimates from color flow data.<br />
Hence, another research topic worth exploring in the future is on developing new imageprocessing-based<br />
flow analysis strategies <strong>for</strong> color flow imaging applications. One possible<br />
approach to this topic is to study whether more robust motion analysis methods like statistical<br />
modeling can be used to analyze color flow data. Such a study may potentially lead to new<br />
insights on how flow detection and flow estimation can be per<strong>for</strong>med in color flow imaging.<br />
5.4 Final Remarks<br />
<strong>Color</strong> flow signal processing has been an active field of research ever since the initial<br />
development of ultrasound flow imaging methods in the 1980s. As an attempt to expand upon<br />
current state of the art in the research field, this thesis has pursued the development of an eigen-
Chapter 5. Thesis Summary and Future Directions 116<br />
based signal processing framework that is based on the properties of Hankel matrix and the<br />
Matrix Pencil <strong>for</strong>mulation. Since color flow imaging is already widely available in commercial<br />
ultrasound systems, it is hoped that the theoretical framework presented in this thesis can<br />
eventually be implemented as a patch modification to the processor architecture of existing<br />
ultrasound systems. As a result of such ef<strong>for</strong>t, the overall reliability of flow in<strong>for</strong>mation<br />
provided in color flow images can potentially be improved.
REFERENCES<br />
Ahn YB, Park SB. Estimation of mean frequency and variance of ultrasonic Doppler signal by<br />
using second-order autoregressive model. IEEE Trans. Ultrason. Ferroelec. Freq. Contr.,<br />
1991; 38: 172-182.<br />
Alam SK, Parker KJ. The butterfly search technique <strong>for</strong> estimation of blood velocity.<br />
<strong>Ultrasound</strong> Med. Biol., 1995; 21: 657-670.<br />
Alam SK, Parker KJ. Implementation issues in ultrasonic flow imaging. <strong>Ultrasound</strong> Med. Biol.,<br />
2003; 29: 517-528.<br />
Allam ME, Greenleaf JF. Isomorphism between pulsed-wave Doppler ultrasound and directionof-arrival<br />
estimation—Part I: basic principles. IEEE Trans. Ultrason. Ferroelec. Freq.<br />
Contr., 1996; 43: 911-922.<br />
Allam ME, Kinnick RR, Greenleaf JF. Isomorphism between pulsed-wave Doppler ultrasound<br />
and direction-of-arrival estimation—Part II: experimental results. IEEE Trans. Ultrason.<br />
Ferroelec. Freq. Contr., 1996; 43: 923-935.<br />
Arbeille P, Bouin-Pineau MH, Herault S. Accuracy of the main Doppler methods <strong>for</strong> evaluating<br />
the degree of carotid stenosis (continuous wave, pulsed wave, and color Doppler).<br />
<strong>Ultrasound</strong> Med. Biol., 1999; 25: 65-73.<br />
Armstrong G, Cardon L, Vilkomerson D, et al. Localization of needle tip with color Doppler<br />
during pericardiocentesis: in vitro validation and initial clinical application. J. Am. Soc.<br />
Echocardiogr., 2001; 14: 29-37.<br />
Badeau R, Richard G, David B. Sliding window adaptive SVD algorithms. IEEE Trans. <strong>Signal</strong><br />
<strong>Processing</strong>, 2004; 52: 1-10.<br />
Beach KW, Keith AC, Primozich JF, et al. Ultrasonic color flow mapping: the visualization of<br />
four-dimensional cardiac and vascular flow phenomena using two dimensions and “realtime”.<br />
<strong>Ultrasound</strong> Med. Biol., 1997; 23: 347-363.<br />
Bjaerum S, Torp H, Kristoffersen K. Clutter filter design <strong>for</strong> ultrasound color flow imaging.<br />
IEEE Trans. Ultrason. Ferroelec. Freq. Contr., 2002a; 49: 204-216.<br />
Bjaerum S, Torp H, Kristoffersen K. Clutter filters adapted to tissue motion in ultrasound color<br />
flow imaging. IEEE Trans. Ultrason. Ferroelec. Freq. Contr., 2002b; 49: 693-704.<br />
Bonnefous O, Pesque P. Time domain <strong>for</strong>mulation of pulsed-Doppler ultrasound and blood<br />
velocity estimation by cross correlation. Ultrason. Imaging, 1986; 8: 73-85.<br />
- 117-
References 118<br />
Brands PJ, Hoeks APG, Reneman RS. The effect of echo suppression on the mean velocity<br />
estimation range of the RF cross-correlation model estimator. <strong>Ultrasound</strong> Med. Biol., 1995;<br />
21: 945-959.<br />
Bruce M, Averkiou M, Tieman K, et al. Vascular flow and perfusion imaging with ultrasound<br />
contrast agents. <strong>Ultrasound</strong> Med. Biol., 2004; 30: 735-743.<br />
Capineri L, Scabia M, Masotti L. A Doppler system <strong>for</strong> dynamic vector velocity maps.<br />
<strong>Ultrasound</strong> Med. Biol., 2002; 28: 237-248.<br />
Chang PH, Shung KK, Levene HB. Quantitative measurements of second harmonic Doppler<br />
using ultrasound contrast agents. <strong>Ultrasound</strong> Med. Biol., 1996; 22: 1205-1214.<br />
Chornoboy ES. Initialization <strong>for</strong> improved IIR filter per<strong>for</strong>mance. IEEE Trans. <strong>Signal</strong><br />
<strong>Processing</strong>, 1992; 40: 543-550.<br />
Christopher DA, Burns PN, Starkoski BG, Foster FS. A high frequency pulsed-wave Doppler<br />
ultrasound system <strong>for</strong> the detection of blood flow in the microcirculation. <strong>Ultrasound</strong> Med.<br />
Biol., 1997; 23: 997-1015.<br />
Cloutier G, Chen D, Durand LG. A new clutter rejection algorithm <strong>for</strong> Doppler ultrasound.<br />
IEEE Trans. Med. Imag., 2003; 22: 530-538.<br />
Cobbold RSC. Foundations of Biomedical <strong>Ultrasound</strong>. New York City, USA: Ox<strong>for</strong>d<br />
University Press, 2006.<br />
del Rio JEF, Sarkar TK. Comparison between the Matrix Pencil method and the Fourier<br />
trans<strong>for</strong>m technique <strong>for</strong> high-resolution spectral estimation. Digital <strong>Signal</strong> Process., 1996; 6:<br />
108-125.<br />
Dunmire B, Beach KW, Labs KH, et al. Cross-beam vector Doppler ultrasound <strong>for</strong> angleindependent<br />
velocity measurements. <strong>Ultrasound</strong> Med. Biol., 2000; 26: 1213-1235.<br />
Embree PM, O’Brien Jr WD. Pulsed Doppler accuracy assessment due to frequency-dependent<br />
attenuation and Rayleigh scattering error sources. IEEE Trans. Biomed. Eng., 1990; 37: 322-<br />
326.<br />
Evans DH, McDicken WN. Doppler <strong>Ultrasound</strong>: Physics, Instrumentation and <strong>Signal</strong><br />
<strong>Processing</strong>. 2 nd Ed. West Sussex, UK: John Wiley & Sons Ltd., 2000.<br />
Fenster A, Downey DB, Cardinal HN. Three-dimensional ultrasound imaging. Physics Med.<br />
Biol., 2001; 46: R66-R99.<br />
Ferrara KW, Algazi VR. A new wideband spread target maximum likelihood estimator <strong>for</strong><br />
blood velocity estimation—part I: theory. IEEE Trans. Ultrason. Ferroelec. Freq. Contr.,<br />
1991a; 38: 1-16.<br />
Ferrara KW, Algazi VR. A new wideband spread target maximum likelihood estimator <strong>for</strong><br />
blood velocity estimation—part II: evaluation of estimators with experimental data. IEEE<br />
Trans. Ultrason. Ferroelec. Freq. Contr., 1991b; 38: 17-26.
References 119<br />
Ferrara KW, Algazi VR, Liu J. The effect of frequency dependent scattering and attenuation on<br />
the estimation of blood velocity using ultrasound. IEEE Trans. Ultrason. Ferroelec. Freq.<br />
Contr., 1992; 39: 754-767.<br />
Ferrara K, DeAngelis G. <strong>Color</strong> flow mapping. <strong>Ultrasound</strong> Med. Biol., 1997; 23: 321-345.<br />
Ferrara KW, Zagar BG, Sokil-Melger JB, et al. Estimation of blood velocity with high<br />
frequency ultrasound. IEEE Trans. Ultrason. Ferroelec. Freq. Contr., 1996; 43: 149-157.<br />
Foster FS, Pavlin CJ, Harasiewicz KA, et al. Advances in ultrasound biomicroscopy.<br />
<strong>Ultrasound</strong> Med. Biol., 2000; 26: 1-27.<br />
Frinking PJA, Bouakaz A, Kirkhorn J, et al. <strong>Ultrasound</strong> contrast imaging: current and new<br />
potential methods. <strong>Ultrasound</strong> Med. Biol., 2000; 26: 965-975.<br />
Gallippi CM, Trahey GE. Adaptive clutter filtering via blind source separation <strong>for</strong> twodimensional<br />
ultrasonic blood velocity measurement. Ultrason. Imaging, 2002; 24: 193-214.<br />
Gallippi CM, Nightingale KR, Trahey GE. BSS-based filtering of physiological and ARFIinduced<br />
tissue and blood motion. <strong>Ultrasound</strong> Med. Biol., 2003; 29: 1583-1592.<br />
Gantmacher FR. The Theory of Matrices (Vol. II). New York City, USA: Chelsea Publishing,<br />
1960.<br />
Golub GH, van Loan CF. Matrix Computations. 3 rd Ed. Baltimore, USA: John Hopkins<br />
University Press, 1996.<br />
Goertz DE, Christopher DA, Yu JL, et al. High-frequency color flow imaging of the<br />
microcirculation. <strong>Ultrasound</strong> Med. Biol., 2000; 26: 63-71.<br />
Hayes MH. Statistical Digital <strong>Signal</strong> <strong>Processing</strong> and Modeling. New York City, USA: John<br />
Wiley & Sons Ltd., 1996.<br />
Haykin S. Communication Systems. 4 th Ed. New York City, USA: John Wiley & Sons Ltd.,<br />
2001.<br />
Heart and Stroke Foundation of Canada. The Growing Burden of Heart Disease and Stroke in<br />
Canada 2003. Ottawa, Canada: 2003.<br />
Heimdal A, Torp H. <strong>Ultrasound</strong> Doppler measurements of low velocity blood flow: limitations<br />
due to clutter signals from vibrating muscles. IEEE Trans. Ultrason. Ferroelec. Freq.<br />
Contr., 1997; 44: 873-881.<br />
Hein IA, O’Brien WD. Current time-domain methods <strong>for</strong> assessing tissue motion by analysis<br />
from reflected ultrasound echoes—a review. IEEE Trans. Ultrason. Ferroelec. Freq. Contr.,<br />
1993; 40: 84-102.<br />
Hoeks APG, van de Vorst JJW, Dabekaussen A, et al. An efficient algorithm to remove low<br />
frequency Doppler signals in digital Doppler systems. Ultrason. Imaging, 1991; 13: 135-<br />
144.
References 120<br />
Hope Simpson D, Chin CT, Burns PN. Pulse inversion Doppler: a new method <strong>for</strong> detecting<br />
nonlinear echoes from microbubble contrast agents. IEEE Trans. Ultrason. Ferroelec. Freq.<br />
Contr., 1999; 46: 372-382.<br />
Hu B, Gosine RG. A new eigenstructure method <strong>for</strong> sinusoidal signal retrieval in white noise:<br />
estimation and pattern recognition. IEEE Trans. <strong>Signal</strong> <strong>Processing</strong>, 1997; 45: 3073-3083.<br />
Hua Y. Estimating two-dimensional frequencies by matrix enhancement and matrix pencil.<br />
IEEE Trans. <strong>Signal</strong> <strong>Processing</strong>, 1992; 40: 2267-2280.<br />
Hua Y, Sarkar TK. Matrix Pencil method <strong>for</strong> estimating parameters of exponentially<br />
damped/undamped sinusoids in noise. IEEE Trans. Acoust. Speech <strong>Signal</strong> Process., 1990;<br />
38: 814-824.<br />
Hua Y, Sarkar TK. On SVD <strong>for</strong> estimating generalized eigenvalues of singular matrix pencils in<br />
noise. IEEE Trans. <strong>Signal</strong> <strong>Processing</strong>, 1991; 39: 892-900.<br />
Jensen JA. Estimation of Blood Velocities Using <strong>Ultrasound</strong>: A <strong>Signal</strong> <strong>Processing</strong> Approach.<br />
Cambridge, USA: Cambridge University Press, 1996.<br />
Jensen JA, Munk P. A new method <strong>for</strong> estimation of velocity vectors. IEEE Trans. Ultrason.<br />
Ferroelec. Freq. Contr., 1998; 45: 837-851.<br />
Kadi AP, Loupas T. On the per<strong>for</strong>mance of regression and step-initialized IIR clutter filters <strong>for</strong><br />
color Doppler systems in diagnostic medical ultrasound. IEEE Trans. Ultrason. Ferroelec.<br />
Freq. Contr., 1995; 42: 927-937.<br />
Kargel C, Hobenreich G, Trummer B, Insana MF. Adaptive clutter rejection filtering in<br />
ultrasonic strain-flow imaging. IEEE Trans. Ultrason. Ferroelec. Freq. Contr., 2003; 50:<br />
824-835.<br />
Kasai C, Namekawa K, Koyano A, Omoto R. Real-time two-dimensional blood flow imaging<br />
using an autocorrelation technique. IEEE Trans. Sonics Ultrason., 1985; 32: 458-464.<br />
Kristoffersen K, Angelsen BAJ. A time-shared ultrasound Doppler measurement and 2-D<br />
imaging system. IEEE Trans. Biomed. Eng., 1988; 35: 285-295.<br />
Kruse DE, Ferrara KW. A new high resolution color flow system using an eigendecompositionbased<br />
adaptive filter <strong>for</strong> clutter rejection. IEEE Trans. Ultrason. Ferroelec. Freq. Contr.,<br />
2002; 49: 1384-1399.<br />
Kung SY, Arun KS, Rao DVB. State space and singular value decomposition based<br />
approximation methods <strong>for</strong> the harmonic retrieval problem. J. Opt. Soc. Am., 1983; 73:<br />
1799-1811.<br />
Lagalla R, Caruso G, Finazzo M. Monitoring treatment response with color and power Doppler.<br />
Euro. J. Radiology, 1998; 28: S149-S156.
References 121<br />
Ledoux LAF, Brands PJ, Hoeks APG. Reduction of the clutter component in Doppler ultrasound<br />
signals based on singular value decomposition: a simulation study. Ultrason. Imaging, 1997;<br />
19: 1-18.<br />
Lovstakken L, Bjaerum S, Kristoffersen K, Torp H. Real-time adaptive clutter rejection filtering<br />
in color flow imaging using power method iterations. IEEE Trans. Ultrason. Ferroelec.<br />
Freq. Contr., 2006; 53: 1597-1608.<br />
Loupas T, McDicken WN. Low-order complex AR models <strong>for</strong> mean and maximum frequency<br />
estimation in the context of Doppler color flow mapping. IEEE Trans. Ultrason. Ferroelec.<br />
Freq. Contr., 1990; 37: 590-601.<br />
Loupas T, Powers JT, Gill RW. An axial velocity estimator <strong>for</strong> ultrasound blood flow imaging,<br />
based on a full evaluation of the Doppler equation by means of a two-dimensional<br />
autocorrelation approach. IEEE Trans. Ultrason. Ferroelec. Freq. Contr., 1995a; 42: 672-<br />
688.<br />
Loupas T, Peterson RB, Gill RW. Experimental evaluation of velocity and power estimation <strong>for</strong><br />
ultrasound blood flow imaging, by means of a two-dimensional autocorrelation approach.<br />
IEEE Trans. Ultrason. Ferroelec. Freq. Contr., 1995b; 42: 689-699.<br />
Maniatis TA, Cobbold RSC, Johnston KW. Two-dimensional velocity reconstruction strategies<br />
<strong>for</strong> color flow Doppler ultrasound images. <strong>Ultrasound</strong> Med. Biol., 1994; 20: 137-145.<br />
Mo LYL, Cobbold RSC. A unified approach to modeling the backscattered Doppler ultrasound<br />
from blood. IEEE Trans. Biomed. Eng., 1992; 39: 450-461.<br />
Mo LYL, Ji TL, Chou CH, et al. Zone-based color flow imaging. Proc. IEEE Ultrason. Symp.,<br />
pp. 29-32, 2003.<br />
Moon TK, Stirling WC. Mathematical methods and algorithms <strong>for</strong> signal processing. Upper<br />
Saddle River, USA: Prentice-Hall Inc., 2000.<br />
Nederkoorn PJ, van der Graaf Y, Hunink MGM. Duplex ultrasound and magnetic resonance<br />
angiography compared with digital subtraction angiography in carotid artery stenosis: a<br />
systematic review. Stroke, 2003; 34: 1324-1332.<br />
Nowicki A, Reid J, Pedersen PC, et al. On the behavior of instantaneous frequency estimators<br />
implemented on Doppler flow imagers. <strong>Ultrasound</strong> Med. Biol., 1990; 16: 511-518.<br />
Oppenheim AV, Schafer RW, Buck JR. Discrete-time signal processing. 2 nd Ed. Upper Saddle<br />
River, USA: Prentice-Hall Inc., 1999.<br />
Overbeck JR, Beach KW, Strandness Jr DE. Vector Doppler: accurate measurement of blood<br />
velocity in two dimensions. <strong>Ultrasound</strong> Med. Biol., 1992; 18: 19-31.<br />
Peterson RB, Atlas LE, Beach KW. A comparison of IIR initialization techniques <strong>for</strong> improved<br />
color Doppler wall filter per<strong>for</strong>mance. Proc. IEEE Ultrason. Symp., pp. 1705-1708, 1994.
References 122<br />
Picot PA, Rickey DW, Mitchell R, et al. Three-dimensional colour Doppler imaging.<br />
<strong>Ultrasound</strong> Med. Biol., 1993; 19: 95-104.<br />
Poon MWY, Khan RH, Le-Ngoc S. A singular value decomposition (SVD) based method <strong>for</strong><br />
suppressing ocean clutter in high frequency radar. IEEE Trans. <strong>Signal</strong> <strong>Processing</strong>, 1993; 41:<br />
1421-1425.<br />
Rajaonah JC, Dousse B, Meister JJ. Compensation of the bias caused by the wall filter on the<br />
mean Doppler frequency. IEEE Trans. Ultrason. Ferroelec. Freq. Contr., 1994; 41: 812-<br />
819.<br />
Round WH, Bates RHT. Modification of spectra of pulses from ultrasonic transducers by<br />
scatterers in non-attenuating and attenuating media. Ultrason. Imaging, 1987; 9: 18-28.<br />
Roy R, Kailath T. ESPRIT—estimation of signal parameters via rotational invariance<br />
techniques. IEEE Trans. Acoust. Speech <strong>Signal</strong> Process., 1989; 37: 984-995.<br />
Schrope BA, Newhouse VL. Second harmonic ultrasonic blood perfusion measurement.<br />
<strong>Ultrasound</strong> Med. Biol., 1993; 19: 567-579.<br />
Stewart SFC. Effects of transducer, velocity, Doppler angle, and instrument settings on the<br />
accuracy of color Doppler ultrasound. <strong>Ultrasound</strong> Med. Biol., 2001; 27: 551-564.<br />
Stoica P, Soderstrom T. Statistical analysis of MUSIC and subspace rotation estimates of<br />
sinusoidal frequencies. IEEE Trans. <strong>Signal</strong> <strong>Processing</strong>, 1991; 39: 1836-1847.<br />
Stoica P, Moses RL. Introduction to spectral analysis. Upper Saddle River, USA: Prentice Hall<br />
Inc., 1997.<br />
Tamura T, Cobbold RSC, Johnston KW. Quantitative study of steady flow using color Doppler<br />
ultrasound. <strong>Ultrasound</strong> Med. Biol., 1991; 17: 595-605.<br />
Thomas L, Hall A. An improved wall filter <strong>for</strong> flow imaging of low velocity flow. Proc. IEEE<br />
Ultrason. Symp., pp. 1701-1704, 1994.<br />
Torp H. Clutter rejection filters in color flow imaging: a theoretical approach. IEEE Trans.<br />
Ultrason. Ferroelec. Freq. Contr., 1997; 44: 417-424.<br />
Torp H, Kristoffersen K, Angelson BAJ. Autocorrelation techniques in color flow imaging:<br />
signal model and statistical properties of the autocorrelation estimates. IEEE Trans.<br />
Ultrason. Ferroelec. Freq. Contr., 1994; 41: 604-612.<br />
Tufts DW, Kumaresan R. Estimation of frequencies of multiple sinusoids: making linear<br />
prediction per<strong>for</strong>m like maximum likelihood. Proc. IEEE, 1982; 70: 975-989.<br />
Tysoe C, Evans DH. Bias in mean frequency estimation of Doppler signals due to wall clutter<br />
filters. <strong>Ultrasound</strong> Med. Biol., 1995; 21: 671-677.<br />
Udesen J, Jensen JA. Investigation of transverse oscillation method. IEEE Trans. Ultrason.<br />
Ferroelec. Freq. Contr., 2006; 53: 959-971.
References 123<br />
Vaitkus PJ, Cobbold RSC. A comparative study and assessment of Doppler ultrasound spectral<br />
estimation techniques. Part I: estimation methods. <strong>Ultrasound</strong> Med. Biol., 1988; 14: 661-<br />
672.<br />
Vaitkus PJ, Cobbold RSC. A new time-domain narrowband velocity estimation technique <strong>for</strong><br />
Doppler ultrasound flow imaging. Part I: theory. IEEE Trans. Ultrason. Ferroelec. Freq.<br />
Contr., 1998; 45: 939-954.<br />
Vaitkus PJ, Cobbold RSC, Johnston KW. A new time-domain narrowband velocity estimation<br />
technique <strong>for</strong> Doppler ultrasound flow imaging Part II: comparative per<strong>for</strong>mance assessment.<br />
IEEE Trans. Ultrason. Ferroelec. Freq. Contr., 1998; 45: 955-971.<br />
van der Veen AJ, Deprettere EF, Swindlehurst AL. Subspace based signal analysis using<br />
singular value decomposition. Proc. IEEE, 1993; 81: 1277-1308.<br />
von Ramm OT, Smith SW, Pavy Jr HG. High-speed ultrasound volumetric imaging system—<br />
Part II: parallel processing and image display. IEEE Trans. Ultrason. Ferroelec. Freq.<br />
Contr., 1991; 38: 109-115.<br />
Waggoner AD, Bierig SM. Tissue Doppler imaging: a useful echocardiographic method <strong>for</strong> the<br />
cardiac sonographer to assess systolic and diastolic ventricular function. J. Am. Soc.<br />
Echocardiogr., 2001; 14: 1143-1152.<br />
Wilkening W, Brendel B, Ermert H. Fast, extended velocity range flow imaging based on<br />
nonuni<strong>for</strong>m sampling using adaptive wall filtering and cross correlation. Proc. IEEE<br />
Ultrason. Symp., pp. 1491-1494, 2002.<br />
Willemetz JC, Nowicki A, Meister JJ, et al. Bias and variance in the estimate of the Doppler<br />
frequency induced by a wall motion filter. Ultrason. Imaging, 1989; 11: 215-225.<br />
Williamson TH, Harris A. <strong>Color</strong> Doppler ultrasound imaging of the eye and orbit. Surv.<br />
Ophthalmol., 1996; 40: 255-267.<br />
Yoo YM, Managuli R, Kim Y. Adaptive clutter filtering <strong>for</strong> ultrasound color flow imaging.<br />
<strong>Ultrasound</strong> Med. Biol., 2003; 29: 1311-1320.<br />
Yu ACH, Cobbold RSC. A Matrix Pencil estimator with adaptive rank selection: application to<br />
in vivo flow studies. Proc. IEEE Ultrason. Symp., 2006, in-press.<br />
Yu ACH, Mo LYL, Warriner RK, Cobbold RSC. A new eigen-based flow estimator using the<br />
Matrix Pencil approach. Proc. IEEE Ultrason. Symp., pp. 1335-1338, 2005.<br />
Yu ACH, Steinman AH, Cobbold RSC. Transit time broadening in pulsed Doppler ultrasound: a<br />
generalized amplitude modulation model. IEEE Trans. Ultrason. Ferroelec. Freq. Contr.,<br />
2006; 53: 530-541.