Multi-resolution statistical analysis on graph structured data in neuroimaging

Statistical data analysis plays a major role in discovering structural and functional imaging phenotypes for mental disorders such as Alzheimer´s disease (AD). The goal here is to identify, ideally early on, which regions in the brain show abnormal variations with a disorder. To make the method more sensitive, we rely on a multi-resolutional perspective of the given data. Since the underlying imaging data (such as cortical surfaces and connectomes) are naturally represented in the form of weighted graphs which lie in a non-Euclidean space, we introduce recent work from the harmonics literature to derive an effective multi-scale descriptor using wavelets on graphs that characterize the local context at each data point. Using this descriptor, we demonstrate experiments where we identify significant differences between AD and control populations using cortical surface data and tractography derived graphs/networks.