Computational analysis and modelling of graph-structured neuroimaging data

Abstract

Graph representations are often used to model structured data at an individual or population level and have numerous applications in pattern recognition problems. This thesis is focusing on the analysis of neuroimaging data that can be intuitively modelled as graphs. In the field of neuroscience, where such representations are commonly used to model structural or functional connectivity between a set of brain regions, graphs have proven to be of great importance. They have contributed to gaining novel insights into the brain's organisation and the mechanisms underlying brain development and disease, which were previously unknown. At an individual level, the underlying parcellation used to construct the connectome, the comprehensive map of neural connections in the brain, and its resolution impact traditional network measures and network-based tasks. This thesis explores how these factors affect topological measures of brain networks and uses the latter to inform predictive models of patient outcome in diseased populations. Additionally, a graph theoretical approach is proposed to establish correspondences between graph elements, when subject-level data-driven parcellation methods are adopted, which is an essential step to perform any further population-level analysis.

The present work also employs concepts from the field of signal processing on graphs and geometric deep learning to address significant problems in disease prediction. More specifically, graph convolutions are adopted for the evaluation of similarity between brain connectivity networks in a manner that accounts for the graph structure and is tailored for classification tasks. At the same time, exploiting the wealth of imaging and non-imaging information for disease prediction tasks requires models capable of simultaneously representing individual features and data associations between subjects from potentially large populations. The latter can be particularly beneficial in large-scale studies and graphs provide a natural framework for such tasks. This work uses geometric deep learning to perform convolutions on a population graph incorporating both imaging and non-imaging information and demonstrates their importance for semi-supervised classification tasks. The proposed framework allows to infer subject-specific properties from their imaging features and interactions within a population.