Marco Battaglini1, Nicola De Stefano1, Mark Jenkinson2
1Neurological and behavioral sciences, University of Siena, Siena, Tuscany, Italy; 2FMRIB Centre, Nuffield Department of Clinical Neurosciences, University of Oxford
A novel hierarchical classification method for segmenting lesions in Multiple Sclerosis is described. It uses two stages: (1) a standard voxel-wise classifier followed by (2) a novel cluster-wise classifier. Features used for the cluster-wise classifier consist of ratios of statistics extracted from within the first-level clusters (across several different image modalities), to those from the exterior borders of the clusters or from all Grey Matter or White Matter voxels. Results on images from a multi-site clinical dataset showed large reductions in False Positives (for voxel-based and lesion-based metrics), with minimal reductions in True Positives, demonstrating great potential for this approach.