A unifying framework for inhomogeneity correction and partial volume segmentation of brain MR images

We propose a unifying framework for fully automated inhomogeneity correction and partial volume (PV) segmentation of multi-spectral brain magnetic resonance (MR) images. The MR data is modeled as a stochastic process with an inherent effect of smoothly varying intensity or bias field. Unlike the conventional hard segmentation methods with a unique label for each voxel, a new PV model is developed in which the percentage of each voxel belonging to each class is considered in establishing the maximum a posteriori (MAP) framework. A new Markov random field (MRF) model is built to reflect the spatial information for the tissue mixture. The MAP solution is calculated by the iterative expectation-maximization (EM) strategy that interleaves PV segmentation with estimations of class parameters and bias field distribution. Experimental studies on clinical MR brain datasets are performed. The results demonstrate that our unifying framework can substantially improve the performance as both bias field and PV effects have been taken into account.