NeuroDetour: A neural pathway transformer for uncovering structural-functional coupling mechanisms in human...

Although modern imaging methods enable in-vivo examination of connections between distinct brain areas, we still lack a comprehensive understanding of how anatomical structure underpins brain function and how spontaneous fluctuations in neural activity give rise to cognition. At the same time, many efforts in machine learning have focused on modeling the complex, nonlinear relationships between neuroimaging signals and observable traits. Yet, current machine learning techniques often overlook fundamental neuroscience insights, making it difficult to interpret transient neural dynamics in terms of cognitive behavior. To bridge this gap, we turn our attention to the interplay between structural connectivity (SC) and functional connectivity (FC), reframing this open question in network neuroscience as a graph representation learning task centered on neural pathways. In particular, we introduce the notion of a "topological detour" to describe how a given instance of FC (i.e., a direct functional connection) is physically supported by underlying SC pathways (the detour), forming a feedback loop between brain structure and function. By considering these multi-hop detour routes that mediate SC-FC coupling, we design a novel multi-head self-attention mechanism within a Transformer architecture. Building on these ideas, we present a biologically inspired deep-learning framework, NeuroDetour, that extracts connectomic feature representations from large-scale neuroimaging datasets and can be applied to downstream tasks such as task classification and disease prediction. We validated NeuroDetour on extensive public cohorts, including the Human Connectome Project (HCP) and UK Biobank (UKB), using both supervised learning and zero-shot settings. In all scenarios, NeuroDetour achieves state-of-the-art results.