Multi-vector Segmentation of Breast MR Images via Hidden Markov Random Fields

In this paper we apply multi-vector Hidden Markov Random Fields to tissue segmentation of Magnetic Resonance (MR) breast images. Our proposed method performs segmentation using a stack of 3 MR breast slices 1mm apart. The approach takes into account neighborhood voxel information rather than merely neighborhood pixel information and the results are anatomically more plausible in comparison with standard two-dimensional segmentation techniques. The proposed algorithm incorporates an initial correction of the bias field, and automatic background removal. The k-means algorithm is used to provide an initial segmentation/classification. This classification allows for tissue parameter estimation, providing an initialization of probabilistic moments that are incorporated into a Gaussian probability model for each tissue class. The class labels follow a Gibbs distribution and the energy function is a sum of potentials taken from a multilevel logistic model for Markov Random Fields. The segmentation is obtained via maximization of the posterior probability distribution function and the solution is found by application of Besag’s Iterated Conditional Modes (ICM) algorithm. After each ICM iteration, the tissue parameters are updated. The process continues iteratively until convergence. The segmentation results demonstrate anatomically plausible breast tissue segmentation and we expect the method to aid real time automatic segmentation of breast tissue, particularly in diagnosis of pathology. Key-Words: Magnetic Resonance Imaging, Breast Imaging, Hidden Markov Random Fields.