Hippocampus Segmentation Based on Orientation-Scale Descriptor and Sparse Coding

In this paper we propose a novel algorithm to increase the accuracy of the hippocampus segmentation by using the orientation-scale descriptor (OSD) and the sparse coding. The orientation-scale descriptor is high dimensional feature which contains image structure information in difference orientations and scales. The segmentation contains four steps. Firstly, extract the orientation-scale descriptors and construct dictionaries from both test image and training images. Secondly, using the dictionaries which are composed of the OSD of the training images to approximately represent the orientation-scale descriptors of the test image. Thirdly fuse the label and coefficients which are obtained from the second step of the dictionary-making voxels. Finally, threshold the fusion value to get the finally label. The similarity between the descriptors obtained from sparse coding using Local Anchor Embedding (LAE) method provide the weight of label fusion process. The probabilistic atlas was generated after the label fusion. Obtain the finally segmentation through a post-processing step based on threshold method. The hippocampus segmentation experiments show a higher precision than other algorithms. The proposed method has been evaluated for brain images on the MICCAI 2013 challenge dataset of 35 subjects. Results show the effectiveness of the proposed method which yields a mean Dice similarity coefficient of 0.8643 and 0.8692 for right hippocampus and left hippocampus, respectively.