Multi-contrast deep nuclei segmentation using a probabilistic atlas

In this paper we propose a new hybrid segmentation approach of the deep brain structures based on a multi-contrast deformable model of regions in competition, with deformations preserving the topology of the structures, as well as their shape and position, using a probabilistic atlas and some prior morphological information. The accuracy of our method was evaluated by comparing the results obtained on a base of T1-weighted data contrast with those of FREESURFER and FSL-FIRST. Besides giving very good results from only one contrast, we show that the multi-contrast aspect of our method allows exploiting the complementary contributions of different contrasts, like T1 and diffusion tensor (DT) contrasts, in order to provide a more robust segmentation.