Breakfast & Science Seminar 54

Abstract

Graph signal processing (GSP) in neuroimaging is the characterization of regional brain activity, measured with functional magnetic resonance imaging (fMRI), as a function of the underlying structure, inferred from diffusion MRI. In GSP, structural connectivity (the quantification of physical wiring between brain regions) is decomposed into a set of connectome harmonics, which are used as a basis over which to express functional signals. This enables to ascertain how much brain structure and function are coupled.

Previous studies utilizing this framework have successfully linked the extent of (de)coupling between brain structure and function to cognitive performance. However, they have also fallen short of outperforming the simpler use of functional connectivity (the statistical interdependence between brain activity time courses, computed without any structural information). In this presentation, I will start by recapitulating the fundamentals of GSP and overview its current applications in neuroimaging, with a focus on behavioural prediction. I will outline several limitations of current analytical workflows that suggest that current GSP approaches have not yet reached their full explanatory potential.

I will then propose ways to overcome these hurdles and test them in the context of behavioural prediction. First, I will touch upon the generation of connectome harmonics. I will show that working with a population-wise average induces biases in behavioural prediction pipelines, and that while more challenging, using subject-wise bases is a more appropriate strategy. Second, I will propose novel ways to extract metrics reflective of brain/structure function coupling, either by the use of more fitting parameters in current approaches, or by the application of a conceptually new framework.