Edge-centric Brain Transformer: An Edge-centric Functional Connectivity Learning Framework for fMRI-based Brain Disorder Diagnosis

arXiv cs.GR details an edge-centric brain transformer (EBT) for resting-state fMRI that shifts the learning unit from brain regions to functional connections. For developers working on medical imaging or graph ML, the key idea is to represent edge time-series directly, then organize those connections in a line graph so the model can learn connection-to-connection relationships instead of only node-centric features.

The framework adds a structure-aware transformer to capture both local dependencies among anatomically related connections and broader interactions across distributed functional networks. It also uses an edge-level orthogonal clustering readout to produce subject-level representations and identify latent connectivity modules tied to disease.

The authors say evaluations across multiple neuroimaging datasets show EBT outperforming representative graph neural networks, brain transformers, and conventional connectivity-based approaches. Just as important, the interpretability results point to stable disease-associated functional connections that line up with known pathological network changes, which matters if you care about explainability rather than just classification accuracy.

The paper was submitted on 20 Sep 2026, and the source code is publicly available. While this is firmly in the medical AI space, the modeling choice is relevant to anyone building graph transformers or trying to preserve edge-level signal that node-centric pipelines tend to smooth away.