Magnetic Resonance Image (MRI) segmentation plays a major role in the tissue quantitative analysis which benefits the early treatment of neurological diseases. In this paper, a new approach to MRI segmentation based on hierarchical Markov random field (MRF) model is proposed: In higher-level MRF, a new mixture model is presented to describe the label image, that is, the interior of region is modeled by homogenous and isotropic MRF while the boundary is modeled by inhomogeneous and anisotropic MRF. So the orientation is incorporated into the boundary information and the characteristic of label image can be more accurately represented. In lower-level MRF, the different Gauss texture is filled in each region to describe pixel image. Then the segmentation problem is formulated as Maximum a Posterior Probability (MAP) estimation rule. A histogram based DAEM algorithm is used, which is able to find the global optima of the standard finite normal mixture (SFNM) parameters. Based on the meaning of prior MRF parameter, an approximate method is proposed to simplify the estimation of those parameters. Experiments on the pathological MRI show that our approach can achieve better results.