Robust unsupervised tissue classification in MR images

A general framework for performing robust, unsupervised tissue classification in magnetic resonance images is presented. Tissue classification is formulated as an estimation problem based on an imaging model. Prior models are used within the estimation problem to compensate for noise and intensity inhomogeneity artifacts. From this framework, approaches based on K-means clustering via the expectation-maximization algorithm, and fuzzy clustering can be derived. The performance of the different types of approaches are evaluated using both simulated and real neuroimaging data.