Abstract: As hippocampus is associated with neurodegenerative diseases,a lot of researchers have proposed many me-thods to segment hippocampus,but the irregularity and the boundary vaguely make the high precise of hippocampus segmentation more difficulty.We proposed a novel algorithm called SCOSD to increase the accuracy of hippocampus segmentation.Motivated by abundant existing dictionary-based methods,SCOSD uses orientation-scale descriptor(OSD) to describe the pixel feature.The OSD contains not only intensity information,such as texture and gradient information,but also the geometrical information.The advantage of OSD is that it reduces the inhomogeneity among different images while containing several low-dimension features.SCOSD method has four steps.Firstly,the orientation-scale descriptors are extracted and dictionaries for each target voxel are constructed.Secondly,corresponding dictionary is used to represent the orientation-scale descriptor of the target voxel approximately and the sparse coefficients can be obtained simultaneous.Thirdly,the label and coefficients of the dictionary are fused to make voxels.Finally,threshold the fusion value to get the finally label.Experiments based on competition data of medical image computing and computer assisted intervention(MICCAI) demonstrate that SCOSD has higher segmentation precise than other algorithms such as Simple,Major Voting,Staple,Collate.

Key words: Hippocampus segmentation, Orientation-Scale descriptor,Spare coding,Hippocampus label fusim,Dictionary