Local variational bayesian inference using niche differential evolution for brain magnetic resonance image segmentation

Abstract

Brain magnetic resonance (MR) image segmentation is pivotal for quantitative brain analyses, in which statistical models are most commonly used. However, in spite of its computational effectiveness, these models are less capable of handling the intensity non-uniformity (INU) and partial volume effect (PVE), and hence may produce less accurate results. In this paper, a novel brain MR image segmentation algorithm is proposed. To address the INU and PVE, voxel values in each small volume are characterized by a local variational Bayes (LVB) model, which is inferred by the niche differential evolution (NDE) technique to avoid local optima. A probabilistic brain atlas is constructed for each image to incorporate the anatomical prior into the segmentation process. The proposed NDE-LVB algorithm has been compared to the variational expectation-maximization based and genetic algorithm based segmentation algorithms and the segmentation routine in the widely used statistical parametric mapping package on both synthetic and clinical brain MR images. Our results suggest that the NDE-LVB algorithm can differentiate major brain tissue types more effectively and produce more accurate segmentation results.