Bayesian tissue segmentation of multispectral brain images

The paper addresses the issue of segmenting simultaneously acquired multispectral magnetic resonance (MR) head scans into tissue classes in a synergetic way. Previous methods have only taken a maximum likelihood approach. We extend the approach by incorporating image priors. Since the independence of T/sub 2/ weighted and proton density (PD) images is not valid, the combined probability of the two images are modeled is a bivariate Gaussian function giving the conditional probability. For the prior model, a multi-level logistic model (MLL) is employed. The maximum a posteriori (MAP) estimate is taken by optimizing the probability of a voxel being a particular tissue type given the corresponding probabilities from both images and the image prior probability. Due to the intractability of minimising the global energy, an iterative suboptimal approach is used instead. Experiments on simultaneously acquired proton density (PD) and T2 weighted images showed encouraging results.