14.08.2013 Views

Docteur de l'université Automatic Segmentation and Shape Analysis ...

Docteur de l'université Automatic Segmentation and Shape Analysis ...

Docteur de l'université Automatic Segmentation and Shape Analysis ...

SHOW MORE
SHOW LESS

You also want an ePaper? Increase the reach of your titles

YUMPU automatically turns print PDFs into web optimized ePapers that Google loves.

Chapter 2 Literature Review 35<br />

2.2.3 Establishment of correspon<strong>de</strong>nce<br />

In or<strong>de</strong>r to build a shape mo<strong>de</strong>l for the l<strong>and</strong>mark data, the correspon<strong>de</strong>nce across<br />

the data set has to be established first. A <strong>de</strong>tailed review of the issue of shape<br />

correspon<strong>de</strong>nce in building the shape mo<strong>de</strong>l has been discussed in the review by<br />

Heimann <strong>and</strong> Meinzer (2009), un<strong>de</strong>r the categories of mesh-to-mesh registration,<br />

mesh-to-volume registration, volume-to-volume registration, parameterization-to-<br />

parameterization registration, <strong>and</strong> population-based optimization. The former<br />

three are registration-based methods without explicitly parameterizing the shape<br />

data, while the latter two are parameterization-based.<br />

The <strong>de</strong>termination of transformation in registration or alignment, <strong>and</strong> the cor-<br />

respon<strong>de</strong>nce problem are closely interrelated. The reciprocity between them is<br />

manifest in the <strong>de</strong>velopment <strong>and</strong> the presentation of the ICP algorithm. The<br />

search for correspon<strong>de</strong>nce is expressed or implied in most registration algorithms.<br />

Pairwise point set registration methods, such as ICP-like closest points (Brett<br />

<strong>and</strong> Taylor, 2000; Vos et al., 2004) <strong>and</strong> non-rigid registration (Subsol et al., 1998;<br />

Fleute <strong>and</strong> Lavallée, 1998; Shelton, 2000; Paulsen <strong>and</strong> Hilger, 2003) are used to<br />

generate the correspon<strong>de</strong>nce. The mo<strong>de</strong>l construction method by Hufnagel et al.<br />

(2009) is based on the EM algorithm, which views the probabilistic assignments<br />

correspon<strong>de</strong>nce as hid<strong>de</strong>n variables. Self-organizing network is used by Ferrarini<br />

et al. (2007) to solve the point correspon<strong>de</strong>nce problem in the shape mo<strong>de</strong>ling <strong>and</strong><br />

analysis.<br />

The <strong>de</strong>formation field produced by volumetric registration between images are<br />

also used to <strong>de</strong>fine the l<strong>and</strong>mark correspon<strong>de</strong>nces (Frangi et al., 2001, 2002). The<br />

l<strong>and</strong>marks are <strong>de</strong>fined by an atlas to which the training images are registered, <strong>and</strong><br />

the l<strong>and</strong>marks are propagated back to the image via the inverse of the <strong>de</strong>forma-<br />

tion from each image to the atlas, resulting in a set of corresponding l<strong>and</strong>marks.<br />

Instead of <strong>de</strong>fining <strong>and</strong> propagating the l<strong>and</strong>marks, the Statistical Deformation<br />

Mo<strong>de</strong>l (SDM, Rueckert et al., 2003) performs the statistical analysis directly on<br />

the <strong>de</strong>formation field from the volumetric registration.

Hooray! Your file is uploaded and ready to be published.

Saved successfully!

Ooh no, something went wrong!