On Detecting Subtle Pathology via Tissue Clustering of Multi-parametric Data using Affinity Propagation

We propose a novel framework for tissue abnormality characterization in normal appearing brain tissue (NABT) that is progressively deteriorating, using affinity propagation applied to multi-parametric data created using a combination of magnetic resonance (MR) protocols. While traditional tissue segmentation and clustering can reveal clusters pertaining to healthy and diseased tissue easily, a complete characterization of the effect of pathology requires the study of heterogeneity of NABT. The problem is rendered challenging by the fact that there are no training samples available for such tissue and hence classification based techniques cannot be used and neither can traditional clustering techniques since the number of clusters are not known a priori. Our framework for the automated clustering of tissue types employs a combination of a) manifold learning, that determines the underlying non-linear structure and embeds it into a lower dimensional space and b) affinity propagation (AP), which is a novel clustering technique that combines model- and similarity- based clustering, to automatically obtain exemplar-based clustering. We also define a novel probabilistic clustering technique. The number of clusters associated with a tissue type is indicative of its heterogeneity. By computing the overlap of these clusters in each of the MR protocols, we obtain a measure of the degree of abnormality in a tissue type and the protocol most sensitive in providing that classification. This general framework is applied towards the characterization of NABT in patients with multiple sclerosis. Results demonstrate a greater heterogeneity in NABT surrounding the lesions along with a greater overlap between the NABT and lesion tissue.