Introducing Edge-Wise Graph Signal Processing: Application to Connectome Fingerprinting

Graph signal processing (GSP) enables the principled study of signals that live on an underlying graph. In magnetic resonance imaging, this graph classically describes structural connectivity between brain regions, while the signals of interest are regional time courses of functional activity. Here, we propose to translate the GSP concepts applied to regional brain data to the edge-centric domain. We create a graph encoding the relationships between structural connections rather than regions, and we study time courses reflective of co-fluctuations between pairs of region. We show that when leveraging GSP tools in this new domain, the extracted features enable to fingerprint individual subjects with an accuracy on par with that of regional GSP coupled to lower between-subject similarity, making edge-wise GSP a more robust alternative relying on distinct structure/function attributes. Our approach opens up new perspectives of application for the versatile GSP toolkit, and provides novel insight into the brain features that make an individual unique.