Multi-atlas segmentation without registration: a supervoxel-based approach.

摘要: Multi-atlas segmentation is a powerful technique. It has two components: label transfer that transfers labels from prelabeled atlases to novel image and fusion combines the results. For reliable transfer, most methods assume structure of interest be segmented have localized spatial support across different subjects. Although technique been successful for many applications, strong assumption also limits its applicability. example, multi-atlas not applied tumor because it difficult derive such applications due substantial variation in locations. To address this limitation, we propose segmentation. Inspired by Superparsing work [13], approach problem steps. Our method first oversegments images into homogeneous regions, called supervoxels. voxel image, find correspondence locate supervoxels are similar supervoxel target belongs to. Then, voxel-wise found through searching voxels patches within selected atlas We apply brain show promising