Automatic Segmentation of 3-D Brain MR Images by Using Global Tissue Spatial Structure Information

Segmentation of brain tissues from MR images is medically valuable for helping to assess many diseases. In this paper, we propose a three-layer Gaussian mixture model framework (3L-GMM) for fully automatic tissue segmentation of three-dimensional brain MR images by using spatial structure information. It uses separate GMMs to model the intensity information, the spatial structure information, and the intensity-spatial feature vector, respectively. We implement the brain tissues segmentation task by maximizing the a posteriori probability of the 3L-GMM model. Experiments are conducted on the three-dimensional, T1-weighted, simulated and in vivo MR images of the BrainWeb and IBSR data sets. The qualitative and quantitative comparisons with the gold standard demonstrate that the proposed model can achieve performance improvement over the state-of-the-art methods in the literature.