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.

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

40 population specific atlases were evaluated in terms of their agreement when<br />

propagated <strong>and</strong> fused to target images. Comparing to the 18 images in the IBSR<br />

atlases, the population specific atlas set built from the el<strong>de</strong>rly population reaches<br />

a higher level of consensus generic IBSR atlases. The result segmentations are<br />

produced by a majority vote rule with higher certainty. Although as a supervised<br />

approach our method requires visual inspection, it is still less time consuming <strong>and</strong><br />

costs less than manually segment images.<br />

LWV label fusion gives better accuracy by utilizing the local image similarity infor-<br />

mation, which makes it possible to improve the segmentation quality by increasing<br />

the number of atlases. It is necessary to select the registration results when the<br />

number of atlas to be fused is limited. As a combination optimization problem,<br />

selection by exhaustive search is not tractable. Conventional heuristics such as<br />

similarity ranking selects the atlases most close to the query image while does<br />

not consi<strong>de</strong>r the inter-atlas redundancy. Selecting atlases re-ranked according to<br />

MMR <strong>and</strong> LAR re-ranking is more efficient compared to image similarity selection<br />

when labels are fused by LWV. They provi<strong>de</strong> more accurate results when the same<br />

number of atlases are selected <strong>and</strong> fused. These methods are advantageous when<br />

the number the atlases to be fused is limited by the computation time, memory<br />

constraint <strong>and</strong>/or the size of atlas set. In the future work, it will be of interest<br />

to search for the selection of atlases based on the combination of current methods<br />

in or<strong>de</strong>r to optimize the performance of the label fusion over varying sizes of the<br />

atlas set.

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

Saved successfully!

Ooh no, something went wrong!