Functional Connectome Fingerprinting with Neural Networks for Structured Data

This thesis aims to explore the capabilities and limitations of Neural Networks for structured data in analyzing functional Magnetic Resonance Imaging (fMRI) data for Connectome Fingerprinting, the task of identifying individuals based on unique brain connectivity patterns. Using the extensive Human Connectome Project's 1200 Subjects Data Release, we convert fMRI scans into graph structures, where nodes represent regions of interest (ROIs) and edges measure functional connectivity between ROIs. Using these graphs, we test a series of static and dynamic models for fingerprinting in different settings using resting-state fMRI scans. Our experiments include using state-of-the-art GNN models for fMRI analysis, namely cGCN and BrainGNN. Alongside identification accuracy, we focus on the interpretability of our results, assessing the influence of the ROIs and the functional networks on fingerprinting performance, and exploring the impact of the temporal and spatial aspects and their interactions in determining individual patterns of connectivity.