Geometric deep learning of brain connectomes over the spectrum of dementia

Alzheimer's Disease Dementia (ADD) is a progressive neurodegenerative disorder that impairs cognitive function with different stages. Accurate diagnosis and monitoring of ADD progression are crucial for early intervention and treatment. In this thesis, we investigate the application of Geometric Deep Learning, specifically Graph Neural Networks (GNNs), to brain network analysis for ADD diagnosis, with a focus on structural and functional brain networks. The thesis is structured around three primary objectives: (1) evaluating GNNs for structural, functional, and multi-modal brain network analysis, (2) developing a novel GNN-based biomarker quantifying structure-function relationship for ADD diagnosis, and (3) proposing a novel GNN that is capable of highlighting ADD-related subnetworks. Through these objectives, we seek to provide both improved diagnostic tools and deeper insights into the progression of ADD. In summary, this thesis demonstrates the power and potential of GNNs in diagnosing and monitoring ADD, presenting state-of-the-art methods that offer not only improved performance but also a neurological basis for the explanations of the results. These advancements aim to enhance both the accuracy and interpretability of ADD diagnostics, ultimately contributing to a better understanding of disease progression and facilitating early intervention strategies.