Assessing selection methods in the context of multi-atlas based segmentation

In atlas-based segmentation, using one single atlas for segmenting all patients introduces a bias. Multi-atlas techniques overcome this drawback by selecting and fusing the most appropriate atlases among a database for a given patient. Globally assessing different multi-atlas strategies provides a biased evaluation of the atlas selection methods. To address this problem, we propose to evaluate atlas selection methods independently from the number of atlases selected and from the atlas fusion step. Briefly, we first cluster the selection methods on the basis of rank correlation and then assess each sub-group of methods with respect to a sub-group of reference selection methods. We apply our method to 105 images of the head and neck region.