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

Create successful ePaper yourself

Turn your PDF publications into a flip-book with our unique Google optimized e-Paper software.

Chapter 3 Hippocampal segmentation using multiple atlases 45<br />

atlases, the result segmentation ˆ L for the image I can be obtained by combining<br />

the transformed label maps<br />

{Lk ◦ Tk : k = 1, 2, · · · , n}. (3.5)<br />

The mis-alignment in the transformed atlases due to the registration error is prop-<br />

agated to the segmentation result, which can be reduced by selecting the atlases<br />

according to the registration accuracy measured by image similarity. In the simi-<br />

larity based atlas selection, the atlases are ranked by the image similarity between<br />

the target image I <strong>and</strong> the registration result Ik◦Tk. The subset A ⊂ {1, 2, · · · , n}<br />

of total n atlases best registered to the target is selected <strong>and</strong> combined in the fu-<br />

sion step. Given the selected atlases, the labeling of the query image is <strong>de</strong>termined<br />

by the consensus of transformed segmentations. Vote rule is a simple but robust<br />

method to produce the consensus segmentation, in which the label for each voxel<br />

ˆL(x) is estimated as the label that accounts for the majority in {Lk(x) : k ∈ A }.<br />

This process of segmentation propagation using multiple atlases is illustrated in<br />

Figure 3.1.<br />

3.1.1 Atlas registration<br />

The transformation T between the source IS <strong>and</strong> the target IT are estimated in two<br />

steps. Firstly, the transform T is restricted to the space of affine transformations,<br />

<strong>and</strong> the two images are matched by rigid <strong>and</strong>/or affine registration algorithms.<br />

Then, after solving the affine component T A of the transform T , NRR algorithms<br />

are used to register the image IS ◦ T A to IT . The final transform T between IS<br />

<strong>and</strong> IT is the composition<br />

T = T A ◦ T N , (3.6)<br />

in which T N is the non-rigid <strong>de</strong>formation found by the NNR algorithm.

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

Saved successfully!

Ooh no, something went wrong!