An a contrario approach for outliers segmentation: Application to Multiple Sclerosis in MRI

The detection of Multiple Sclerosis (MS) lesions in Magnetic Resonance (MR) images remains an important issue in medical image processing. Diagnostic criteria for MS based on brain MRI concern mainly dissemination in space and time. In this context, this paper describes a novel region- based approach to automatically count the number of MS lesions present in a set of MR images. Given a set of candidate regions obtained with a mean-shift based segmentation, the detection algorithm decides for each region if it is part of a MS lesion or if it belongs to non-pathologic regions (white matter (WM), grey matter (GM) or cerebro-spinal fluid (CSF)). The distribution of each brain tissue is modeled using a Gaussian Mixture Model and MS lesions are detected as outliers with respect to this model. Finally, we propose several criteria for segmentation assessment and we validate our algorithm on the Brain Web data set. Preliminary results on clinical data are also shown.