Segmentation of brain MR images using hidden Markov random field model with weighting neighborhood system

Current state-of-the-art segmentation techniques of brain MR images improve segmentation accuracy by encoding spatial information through hidden Markov random field (HMRF) model. However, HMRF model has higher computational overhead compared to finite Gaussian mixture (FGM) model but the segmentation results are with no significant difference when applying to cleaner data. We believe this is because the spatial constraint is too simple to utilize the characteristics of the brain. In this paper, we propose a novel method to improve the neighborhood system of the HMRF model by better characterizing natural structures of human brain. Experiments on both real and synthetic 3D brain MR images show that the segmentation results of our method have higher accuracy compared to existing solutions.