Co-segmentation of inter-subject brain magnetic resonance images
Jongseong Jang, Hyung Wook Kim, Young Soo Kim · 2014
In this paper, a co-segmentation method to extract the cortex in inter-subject brain MR (Magnetic Resonance) images is proposed. Co-segmentation is a method to segment two images simultaneously. The method employs the MRF (Markov Random Field) based graph for constructing the objective function and the graph-cut algorithm for optimization. In the graph construction, similarity nodes are added to represent similarity between voxels in each image. Voxel intensity and gradient difference are used to calculate similarity. For selection of similar voxel pairs, two volumes are aligned by using the transformation matrix calculated by matching 3 D SIFT features. Additionally, to get moderate number of similar pairs, a search area in the aligned image is limited to 10 × 10 × 10 neighboring voxels. For experiments, a pre-segmented cortex image and a brain image which are segmented are used as a reference and a target image, respectively. The method showed moderate performance, however, a lack for representing the complex region of interest should be resolved. To improve details, parameter optimization is required. As a further study, other applications, such as multi-modality volume segmentation, are going to be researched.