Adaptive Markov modeling for mutual-information-based, unsupervised MRI brain-tissue classification

This paper presents a novel method for brain-tissue classification in magnetic resonance (MR) images that relies on a very general, adaptive statistical model of image neighborhoods. The method models MR-tissue intensities as derived from stationary random fields. It models the associated Markov statistics nonparametrically via a data-driven strategy. This paper describes the essential theoretical aspects underpinning adaptive, nonparametric Markov modeling and the theory behind the consistency of such a model. This general formulation enables the method to easily adapt to various kinds of MR images and the associated acquisition artifacts. It implicitly accounts for the intensity nonuniformity and performs reasonably well on T1-weighted MR data without nonuniformity correction. The method minimizes an information-theoretic metric on the probability density functions associated with image neighborhoods to produce an optimal classification. It automatically tunes its important internal parameters based on the information content of the data. Combined with an atlas-based initialization, it is completely automatic. Experiments on real, simulated, and multimodal data demonstrate the advantages of the method over the current state-of-the-art.