Bayesian multiple sclerosis lesion classification modeling regional and local spatial information

This thesis presents a fully automatic Bayesian method for multiple sclerosis lesion classification. Traditionally, human experts locate lesions, which are diseased tissue, on magnetic resonance images (MRI). However, manual classification methods are particularly subjective, as experts locate lesions differently, particularly around the borders of these structures. The proposed approach classifies voxels from MRIs into regular tissue and lesions, thus allowing for an objective and consistent way to locate lesions in order to help track their size and count. Previous automatic classification approaches do not model the variation of the MRI tissue intensities in the brain, so as to accurately locate lesions in the posterior fossa, where the intensities vary significantly from the rest of the brain. To this end, the posterior probability distribution is used to determine MRI voxel labels for background, cerebrospinal fluid, grey matter, white matter, as well as labels for two lesion types which differ due to their appearance on MRIs: T1-hypointense lesions (also called black holes) and T2-hyperintense lesions excluding black holes. Furthermore, the proposed method provides neuropathology experts with a confidence level in the classification, which has not been provided in previous work. Spatial variability in intensity distributions over the brain is explicitly modeled by (1) segmenting the brain into distinct anatomical regions, (2) building the likelihood distributions of each tissue class in each region and (3) modeling each distribution as a multidimensional Gaussian using intensities from multimodal MRIs. Local smoothness is enforced by incorporating Markov random fields in the prior probability and thus taking into account neighboring voxel tissue classes. Qualitative and quantitative validation is performed for both lesion classes on real data from 10 patients with multiple sclerosis. Validation on ten patients for both lesion types has not been performed by previous works. Lesion classification results are compared to classifications performed by several experts and two other automatic classification techniques, using volume count and overlap. Automatic classification results are comparable to manual classifications, thus providing a more consistent and time effective alternative to manual classification. In addition, the proposed method has the advantage of providing a more accurate classification in the posterior fossa, which is a region of the brain that is difficult to classify, and where no other automatic method reports success.