Comparing Graph Neural Networks for Single and Multi-Layer Brain Connectivity Analysis in Multiple Sclerosis
Abstract: Brain network analysis increasingly relies on the integration of multiple data modalities, such as structural connectivity, resting-state functional connectivity, and gray matter morphology, which must be effectively fused to support graph-based learning. In this work, we present a comparative study of various methodologies for processing brain network data with the aim of classifying subjects using Graph Neural Networks (GNNs). Specifically, we explore different strategies for constructing informative graph representations to distinguish people with multiple sclerosis (MS) from healthy controls. Our approach involves both the design of graph topologies and the derivation of node embeddings from multimodal brain features. The use of GNNs in this context requires tailored adaptations to address domain-specific challenges, including the limited receptive field of graph architectures and the graph expansion problem that arises during message passing in densely connected networks.
keywords: Graph Neural Networks, Multi-layer networks ,Multiple Sclerosis, Brain networks, Graph theory