Hippocampus segmentation by optimizing the local contribution of image and prior terms, through graph cuts and multi-atlas

This paper presents a new method for segmentation of ambiguously defined structures, such as the hippocampus, by exploiting prior knowledge from another perspective. An expert's experience of where to use prior knowledge and where image information, is captured as a local weighting map. This map can be used to locally guide the evolution in a level set evolution framework. Such a map is produced for every training image using Graph-cuts to calculate the most suited balance of current and prior information. Training maps are optimally adapted on the test image, through non-rigid registration, producing the Optimum Local Weighting map, which is anatomically the most suitable to this test image. Experimental results demonstrate the efficacy and accuracy of the proposed method.