Application of independent component analysis (ICA) to identify and ...
Application of independent component analysis (ICA) to identify and ...
Application of independent component analysis (ICA) to identify and ...
Create successful ePaper yourself
Turn your PDF publications into a flip-book with our unique Google optimized e-Paper software.
<strong>Application</strong> <strong>of</strong> <strong>independent</strong> <strong>component</strong> <strong>analysis</strong><br />
(<strong>ICA</strong>) <strong>to</strong> <strong>identify</strong> <strong>and</strong> separate tumor arterial<br />
input function (AIF) in dynamic contrast<br />
enhanced-MRI<br />
Hatef Mehrabian<br />
Chaitanya Ch<strong>and</strong>rana, Ian Pang, Rajiv Chopra, Anne Martel<br />
Field MITACS conference<br />
June 20, 2011
Illustrations: courtesy <strong>of</strong> http://www.reishiscience.com/Benefits_2.html<br />
In-vivo images: D. Fukumura et al., MICROVASCULAR RESEARCH, 2007.<br />
Tumor Angiogenesis<br />
100µm
100µm<br />
100µm<br />
Illustrations: courtesy <strong>of</strong> http://www.reishiscience.com/Benefits_2.html<br />
In-vivo images: D. Fukumura et al., MICROVASCULAR RESEARCH, 2007.<br />
Tumor Vasculature<br />
100µm
100µm<br />
100µm<br />
Illustrations: courtesy <strong>of</strong> http://www.reishiscience.com/Benefits_2.html<br />
In-vivo images: D. Fukumura et al., MICROVASCULAR RESEARCH, 2007.<br />
Tumor Vasculature
100µm<br />
Pharmacokinetic Modeling<br />
.. ..<br />
..<br />
.<br />
. ...<br />
.. .<br />
. .<br />
.. .<br />
.
100µm<br />
Pharmacokinetic Modeling
100µm<br />
K trans K ep<br />
Pharmacokinetic Modeling
100µm<br />
K trans K ep<br />
Pharmacokinetic Modeling<br />
Renal Cell Carcinoma<br />
DCE – MRI
MR Pixel<br />
0.5mm<br />
100µm<br />
0.02 mm<br />
K trans K ep<br />
Pharmacokinetic Modeling<br />
MR Slice<br />
0.5mm
IV<br />
EV<br />
Dynamic Contrast Enhanced MRI<br />
X
IV<br />
EV<br />
Dynamic Contrast Enhanced MRI<br />
X
x <br />
s IV<br />
s EV<br />
Dynamic Contrast Enhanced MRI<br />
x
x aIV <br />
s IV<br />
s EV<br />
a: contrast Concentration<br />
Dynamic Contrast Enhanced MRI<br />
x
a: contrast Concentration<br />
s: image<br />
s IV<br />
s EV<br />
x aIV sIV <br />
Dynamic Contrast Enhanced MRI<br />
x
a: contrast Concentration<br />
s: image<br />
s IV<br />
s EV<br />
x aIV IV<br />
s a <br />
sE<br />
Dynamic Contrast Enhanced MRI<br />
EV V<br />
x
x aIV IV<br />
a: contrast Concentration<br />
s: image<br />
s a <br />
sE<br />
Dynamic Contrast Enhanced MRI<br />
EV V<br />
x
X<br />
t<br />
Independent Component Analysis<br />
= +<br />
s s<br />
IV<br />
EV<br />
t
X<br />
t<br />
Independent Component Analysis<br />
= +<br />
s s<br />
IV<br />
EV<br />
X AS<br />
1, 2,...,<br />
N <br />
IV , EV <br />
<br />
s , <br />
X x x x<br />
A a a<br />
S s<br />
IV EV<br />
T<br />
t
Estima<strong>to</strong>r:<br />
X<br />
F A<br />
X | s , s <br />
t<br />
IV EV<br />
= +<br />
s s<br />
IV<br />
EV<br />
t<br />
Estimation Model
Estima<strong>to</strong>r: F X | A<br />
s , s <br />
X<br />
t<br />
Maximum Likelihood Estimation<br />
IV EV<br />
Maximum likelihood estima<strong>to</strong>r corresponds <strong>to</strong> the value<br />
θ ML that makes the obtained measurements more likely.<br />
<br />
<br />
1 2<br />
ML X | arg max{ p( x , x ,..., x )}<br />
= +<br />
<br />
s s<br />
IV<br />
EV<br />
t<br />
N
Estima<strong>to</strong>r:<br />
F A<br />
t<br />
Maximum Likelihood Estimation<br />
X | s , s <br />
IV EV<br />
= +<br />
T <br />
1 2 N<br />
ML X | A<br />
arg max px, x ,..., x | A <br />
<br />
X<br />
A<br />
s s<br />
IV<br />
EV<br />
t
Estima<strong>to</strong>r:<br />
F A<br />
t<br />
Maximum Likelihood Estimation<br />
X | s , s <br />
IV EV<br />
= +<br />
T <br />
1 2 N<br />
ML X | A<br />
arg max px, x ,..., x | A <br />
<br />
X<br />
A<br />
X AS<br />
p X p AS A p S<br />
1<br />
( ) ( ) det( ) <br />
( )<br />
s s<br />
IV<br />
EV<br />
t
Estima<strong>to</strong>r:<br />
F A<br />
t<br />
Maximum Likelihood Estimation<br />
X | s , s <br />
IV EV<br />
= +<br />
T <br />
1 2 N<br />
ML X | A<br />
arg max <br />
<br />
p x , x ,..., x<br />
A<br />
1<br />
argmax det A p S | A<br />
<br />
| A <br />
<br />
X<br />
A<br />
s s<br />
IV<br />
EV<br />
t
Estima<strong>to</strong>r:<br />
F A<br />
t<br />
Maximum Likelihood Estimation<br />
X | s , s <br />
IV EV<br />
= +<br />
T <br />
1 2 N<br />
ML X | A<br />
arg max <br />
<br />
p x , x ,..., x<br />
A<br />
| A <br />
<br />
1<br />
argmax det A p S | A<br />
A <br />
1<br />
argma x detA p s , | <br />
IV sEV A <br />
<br />
X<br />
A<br />
s s<br />
IV<br />
EV<br />
t
Estima<strong>to</strong>r:<br />
F A<br />
t<br />
Maximum Likelihood Estimation<br />
X | s , s <br />
IV EV<br />
= +<br />
T <br />
1 2 N<br />
ML X | A<br />
arg max <br />
<br />
p x , x ,..., x<br />
A<br />
| A <br />
<br />
1<br />
argmax det A p S | A<br />
If s A 1& s2 are <strong>independent</strong> <br />
1<br />
argma x detA p s , | <br />
IV sEV A <br />
p( s , s ) p( s ) <br />
p( s ) <br />
X<br />
A<br />
1 2 1 2<br />
s s<br />
IV<br />
EV<br />
t
Estima<strong>to</strong>r:<br />
F A<br />
t<br />
Maximum Likelihood Estimation<br />
X | s , s <br />
IV EV<br />
= +<br />
T <br />
1 2 N<br />
ML X | A<br />
arg max <br />
<br />
p x , x ,..., x<br />
A<br />
| A <br />
<br />
1<br />
argmax det A p S | A<br />
A <br />
1<br />
argma x detA p s , | <br />
IV sEV A <br />
A <br />
1<br />
arg mx a det A p s IV | ApsEV | A<br />
<br />
X<br />
A<br />
s s<br />
IV<br />
EV<br />
t
Independent Component Analysis<br />
<br />
1<br />
ML X | A arg max de t A p s | | <br />
IV A p sEV<br />
A <br />
<br />
A
Independent Component Analysis<br />
<br />
1<br />
ML X | A arg max de t A p s | | <br />
IV A p sEV<br />
A <br />
<br />
A<br />
p s ? i
s , s s ps p p<br />
Probability Density Function (pdf)<br />
i j i j<br />
0<br />
Es 2<br />
p s <br />
? i<br />
i<br />
<br />
2<br />
si<br />
E1
Probability Density Function (pdf)<br />
p s <br />
? i<br />
s 2<br />
p s <br />
p <br />
E s<br />
i i<br />
<br />
i<br />
s s<br />
2<br />
<br />
i i <br />
0
Independent Component Analysis (<strong>ICA</strong>)<br />
<br />
1<br />
ML X | A arg max de t A p s | | <br />
IV A p sEV<br />
A <br />
<br />
A<br />
<br />
1<br />
2log cosh <br />
exp <br />
p s <br />
i i s
• Dialysis Tubing<br />
2 cm<br />
2.5 cm<br />
Phan<strong>to</strong>m Study
• Dialysis Tubing<br />
• Pore size: 90-970 nm<br />
2 cm<br />
2.5 cm<br />
Phan<strong>to</strong>m Study<br />
300 µm
Phan<strong>to</strong>m Study<br />
1 mm<br />
Raw Data<br />
in-plane resolution=300µm
Intravascular<br />
Extravascular<br />
1 mm 1 mm 1 mm<br />
Separation Results (<strong>ICA</strong>)<br />
Phan<strong>to</strong>m Study<br />
Raw Data
• High resolution pre-contrast MR image<br />
1 mm 1 mm<br />
Validation<br />
DCE MRI Dataset High Resolution (Pre-contrast)
• High resolution pre-contrast MR image<br />
– Thresholding Binary mask<br />
1 mm 1 mm<br />
DCE MRI Dataset Mask<br />
Validation
Intravascular<br />
Extravascular<br />
1 mm 1 mm 1 mm<br />
Phan<strong>to</strong>m Study<br />
Raw Data<br />
Separation Results (<strong>ICA</strong>) Separation Results (Mask)
• Tumor in the Thigh muscle<br />
• Imaged with US <strong>and</strong> MRI<br />
In-vivo Experiment (MR data)<br />
1 cm<br />
Raw Data
• Tumor in the Thigh muscle<br />
• Imaged with US <strong>and</strong> MRI<br />
In-vivo Experiment (MR data)<br />
1 cm<br />
Raw Data<br />
1 cm 1 cm<br />
Extravascular Intravascular
• Tumor in the Thigh muscle<br />
• Imaged with US <strong>and</strong> MRI<br />
Ultrasound Contrast Agent (µBubbles)<br />
1 cm<br />
Raw Data
• Tumor in the Thigh muscle<br />
• Imaged with US <strong>and</strong> MRI<br />
Circulating µbubbles<br />
Ultrasound Contrast Agent (µBubbles)<br />
1 cm<br />
Raw Data
• Ultrasound Imaging<br />
• µBubbles stay intravascular<br />
In-vivo Experiment (MRI vs. US)<br />
1 cm<br />
1 cm<br />
Raw Data- US<br />
Intravascular-MR
Summary & Conclusions<br />
• Tumor vasculature is heterogeneous <strong>and</strong> leaky<br />
• MRI contrast agent is capable <strong>of</strong> leaking<br />
• Tumor is characterized based on the tracer dynamics<br />
• <strong>ICA</strong> is capable <strong>of</strong> separating the intravascular <strong>and</strong><br />
extravascular compartments <strong>and</strong> <strong>identify</strong>ing arterial<br />
input function<br />
• Results <strong>of</strong> phan<strong>to</strong>m <strong>and</strong> in-vivo experiment studies<br />
were promising
Acknowledgements
?<br />
Thank You