Unsupervised Regularized Classification of Multi-Spectral MRI

In this paper, we present a framework for automatic classification of multi-spectral MRI data. We propose a novel clustering method using a contextual hypothesis which succeeds in discriminating largely overlapping component distributions. The initial classifications obtained for each channel in a multi-spectral study independently are subsequently merged by minimizing a Minimum Description Length (MDL) criterion trading the data-fit accuracy for simplicity of the model (number of classes). In a final stage, we use a refined Markovian prior to regularize the final segmentation. This prior preserves fine structures and linear shapes as opposed to the typically used Ising or Potts MRF priors. This work represents work-in-progress. Results on a limited number of data are presented.