Dynamic Functional Connectome Fingerprinting with Deep Learning for Temporal Graphs

This thesis examines the methodological aspects of dynamic functional connectivity analysis in fMRI data, with particular emphasis on connectome fingerprinting - the process of identifying individuals through their distinctive brain connection patterns. Utilizing the comprehensive Human Connectome Project's 1200 Subjects Data Release, we investigate the impact of various methodological choices on the development and analysis of dynamic brain networks. We methodically assess multiple aspects in the transformation of fMRI scans into dynamic graph topologies, wherein nodes denote areas of interest (ROIs) and edges represent functional connectivity among these regions. We examine the effects of varying temporal window sizes (30s, 40s, and 60s), analyze diverse methodologies for establishing edge weights from correlation matrices (such as absolute thresholding and segregation of positive and negative correlations), and evaluate the influence of different node neighborhood sizes (top 10 and top 20 connections) on network structure and classification efficacy. Following the determination of appropriate parameters by connectome fingerprinting on resting-state fMRI data, we broaden our research to the ABIDE dataset for the categorization of autism. The extension includes a comprehensive exploratory data analysis aimed at identifying regions of interest that demonstrate significant alterations in connection patterns between autism and control groups across various temporal frames.