Research on the Segmentation of Brain MR Images Based on Markov Multiple Feature Random Fields Model
|Keywords||Magnetic resonance imaging Brain magnetic resonance images Image segmentation Feature random field Markov random field|
In recent years, segmentation of brain magnetic resonance (MR) images becomesan important research issue. However, due to various factors appearing on brain MRimages such as intensity inhomogeneities and noise, segmentation remains the difficultyof the current research. In the existing segmentation methods, the Markov random field(MRF) model based algorithms giving reasonable segmentation results, have attractedmore and more experts and scholars of the world to its in-depth study. In this paper, wefocus on the MRF theory and its application to the brain MR image segmentation. Themain results and contributions are as follows:1) We study the MRF model theory, define the concept of multiple feature randomfields (MFRFs) and propose a novel MRF model based on the MFRFs named asMarkov multiple feature random fields (MMFRF) model, which effectively integrated avariety of feature information in the objects and combined them into a unified decisionmodel using the Bayesian framework. And then, we construct the correspondingestimate criteria by maximum a posteriori (MAP) and the optimization solving methodusing iterated conditional modes (ICM). Particularly, the traditional MRF modelbecomes the special case of our MMFRF model when the number of feature randomfields is one (n=1). Furthermore, as a very general method, the developed MMFRFmodel presents a unified probabilistic framework for solving a series of image andvision analysis applications.2) We study on the application of brain MR image segmentation based on theproposed MMFRF model. We investigate the local MMFRF models defining on brainsub-volumes, and through which to construct a MMFRF-based compound model for thewhole brain image domain (MMFRF-SVPA). And then, we present the theoreticalconcepts and practical computations of intensity feature random field, texture featurerandom field, shape feature random field, and label random field, respectively.3) We conduct the brain MR segmentation experiments based on MMFRF-SVPAframework. We design three segmentation algorithms based on the MMFRF-SVPAframework when the number of feature random fields is2and3for comparisons. Theproposed algorithms are validated on simulated and real brain MR data sets bycomparing the competing segmentation methods in current research using multipleevaluation criteria. The quantitatively results and visual comparisons show that the developed algorithms outperform the state-of-the-art methods in terms of robustness toa variety of artifacts such as noise, intensity inhomogeneities, partial volume effect, etc.