A new stochastic model-based image segmentation technique for MR image

A new framework for stochastic MR image modeling is presented based on local randomization, and a new model-based MR image analysis technique is developed in terms of stochastic regularization. By discussing the statistical properties of both pixel image and context image, an inhomogeneous hidden Markov random field model is proposed and justified for MR image. The solution of the new problem formulation is implemented with an efficient multistage procedure. The number of image regions is detected by a new information criterion. The parameters in the model are estimated from block-wise classification-maximization and Bayesian maximum likelihood algorithms. The image segmentation is performed via a maximum a posteriori probability classifier. The experimental results with simulated and real MR images are provided to demonstrate the promise and effectiveness of the proposed technique