A modified MRF segmentation of brain MR images

Abstract

Image noise seriously affects the accuracy of brain magnetic resonance (MR) image classification. The traditional Gaussian mixture model (GMM) is widely applied in image segmentation, which uses only image intensity information thus is susceptible to image noise. Markov random field (MRF) model overcome the problem by using a priori probability for local smoothing and denoising, which depends on the pixel class labelling with the Gibbs distribution. However, the MRF model still lacks accurate classification of brain tissue. To improve the segmentation of brain MR image, the paper proposes a joint segmentation scheme based on combination of GMM and MRF model to classify the brain tissue into three parts: white matter (WM), grey matter (GM) and cerebrospinal fluid (CSF). During the implementation, both image intensity and local neighbouring voxels are considered. The main steps are as follows: 1) we use K-means estimation to acquire the initial distribution parameters of the three brain tissues; then use Expectation maximization (EM) estimation to obtain GMM parameters. 2) For the spatial voxels, the energy function of joint probability is change in terms of their corresponding GMM class information; then the intensity-related energy function's component is controlled. Therefore, the parameter of the joint model is estimated adaptively according to the local intensity and spatial information. The proposed method performs well in the experiments, for example, it is not sensitive to image noise, and has good robustness and segmentation accuracy as well as high computational efficiency.