Probabilistic hyperstack segmentation of MR brain data

Abstract

A multiscale method (the hyperstack) is proposed to segment multidimensional MR brain data. Hyperstack segmentation is based upon the linking of voxels at adjacent levels in scale space, followed by a root selection to find the voxels that represent the segments in the original image. This paper addresses an advanced linking and root labeling method for the hyperstack. In particular, attention will be paid to an extension of the linking scheme for the detection and classification of partial volume voxels. The result—a list of probabilities for each partial volume voxel—improves the resulting segmentations.