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 2 Literature Review 27<br />

In the registration step, commonly used NRR framework based on maximization of<br />

context-free image similarity measure does not ensure correct correspon<strong>de</strong>nces be-<br />

tween the image-pair (Crum et al., 2003). The mismatch of features is propagated<br />

into the result, leading to segmentation errors. In the comparison by Carmichael<br />

et al. (2005), NRR methods (Friston et al., 1995; Woods et al., 1998; Chen, 1999)<br />

fit better to the complex shape of hippocampus than using only affine transfor-<br />

mations (Friston et al., 1995; Woods et al., 1998; Jenkinson et al., 2002). The use<br />

of multiple atlases with ANIMAL techniques has been <strong>de</strong>scribed by Collins <strong>and</strong><br />

Pruessner (2010). Yassa <strong>and</strong> Stark (2009) evaluated current registration methods,<br />

including Talairach alignment <strong>and</strong> 3dWarpDrive in AFNI (Cox, 1996), normal-<br />

ization in SPM (Ashburner <strong>and</strong> Friston, 1999, 2005), LDDMM (Beg et al., 2005;<br />

Miller et al., 2005), Diffeomorphic Anatomical Registration Through Exponen-<br />

tial Lie Algebra (DARTEL, Ashburner, 2007), <strong>and</strong> Demons (Thirion, 1998; Ver-<br />

cauteren et al., 2009), in terms of their performance in aligning the structures in<br />

the medial temporal lobe. The Demons algorithm was found to perform the best<br />

in aligning the hippocampus.<br />

The issue of multi-atlas selection has been discussed by Aljabar et al. (2007, 2009),<br />

in which the similarity based atlas selection strategies are classified into three cat-<br />

egories, namely segmentation similarity, image similarity, <strong>and</strong> <strong>de</strong>mographics, with<br />

the selection by image similarity ranking being the most popular in applications.<br />

The overlap between the segmentation result <strong>and</strong> the ground truth in multi-atlas<br />

segmentation, measured by DSC, reaches its peak when approximately 10 to 20<br />

similarity ranked atlases are selected, <strong>and</strong> combined by a vote rule (Aljabar et al.,<br />

2007), as opposed to the convergence of the DSC, as the number of the atlases<br />

r<strong>and</strong>omly selected increases. When less similar atlases are fused to the segmen-<br />

tation result, the information relevant to the segmentation is outweighed by the<br />

misalignment propagated to the result.<br />

With regard to the label fusion, the basic method is to combine the transformed<br />

label maps by voxel-wise majority voting, <strong>and</strong> the final labeling of the voxel is

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

Saved successfully!

Ooh no, something went wrong!