Brain MRI image segmentation based on learning local variational Gaussian mixture models

Measuring the distribution of major brain tissues, including the gray matter, white matter and cerebrospinal fluid (CSF), using magnetic resonance imaging (MRI) has attracted extensive research efforts. Many brain MRI image segmentation methods in the literature are based on the Gaussian mixture model (GMM), which however is not strictly followed due to the intrinsic complex nature of MRI data and may lead to less accurate results. In this paper, we introduce the variational Bayes inference to brain MRI image segmentation, and thus propose a novel segmentation algorithm based on learning a cohort of local variational Gaussian mixture (LVGM) models. By assuming all Gaussian parameters to be random variables, the LVGM model has more flexibility than GMM in characterizing the complexity of brain voxel distributions. To alleviate the impact of bias field, we train each LVGM model on a sampled small data volume and linearly combine the trained models to classify each brain voxel. We also construct a co-registered probabilistic brain atlas for each MRI image to incorporate the prior knowledge about brain anatomy into the segmentation process. The proposed LVGM learning algorithm has been evaluated against five state-of-the-art brain MRI image segmentation methods on both synthetic and clinical data. Our results suggest that the LVGM algorithm can segment brain MRI images more effectively and provide more precise distribution of major brain tissues.