MS Lesion Segmentation using Markov Random Fields

We present a fully automated framework for identifying multiple sclerosis (MS) lesions from multispectral human brain magnetic resonance images (MRIs). The brain tissue intensities and lesions are both modeled using Markov Random Fields (MRFs) to incorporate local spatial variations and neighborhood information. In this work, we model all brain tissues, including lesions, as separate classes as opposed to the common approach of modelling the lesions as outliers of the brain tissues. A maximum probability estimate is obtained by arriving at the global convergence of the MRFs using Simulated Annealing. Finally, probability surface discontinuities due to noise and local intensity variations are avoided by incorporating a spline based smoothing function following the MRF modelling. The algorithm is validated on a set of real MRI brain volumes of MS patients with widely varying lesion loads by comparing the results against a silver standard derived from manual expert labellings. The algorithm yields favorable results, including in the posterior fossa where few methods have proved successful. Further, our algorithm yields fewer false negatives than is usual in practice.