Atlas Matters: Edge Quadratics for Consistent Brain Connectivity Prediction

Sai Karthik Navuluru, Surbhi Kumar, Lakshman Tamil, Baris Coskunuzer

Keywords: brain functional connectivity, neuroimaging, quadratic networks, CNNs, brain networks, ASPP, Low-rank quadratic interactions, Edge-image modeling, benchmark

TL;DR: dgeQuad encodes fMRI connectivity as an edge image: dual ASPP + low-rank quadratics capture multi-scale edge interactions. In a unified cross-atlas protocol, it outperforms recent SOTA (ABIDE/ADNI) while staying consistent, efficient, and robust.

Abstract:

Functional connectivity from resting-state fMRI is a strong substrate for subject-level prediction, yet progress is held back by two issues. First, most architectures ingest FC via node-centric propagation or global attention, leaving higher-order edge interactions implicit. Second, evaluations are inconsistent across seeds, atlas choice, preprocessing, and hyperparameter budgets, which obscures true gains.

We propose a simple edge-image encoder that applies dual atrous spatial pyramid pooling to features and connectivity, coupled with a low-rank quadratic block that makes edge-edge effects explicit and efficient. Beyond design, we introduce a unified protocol with five fixed seeds, harmonized preprocessing, and multiple standard atlases, and we re-run recent GNN and transformer baselines under identical settings. Under this protocol, our model **EdgeQuad** attains the best mean performance on curated functional atlases for ABIDE and ADNI, while on unsupervised parcellations such as Ward and KMeans rankings are mixed, highlighting sensitivity to atlas construction. The quadratic block realizes localized degree-2 interactions with provable stability, explaining robustness. The model is lightweight and computationally efficient. To facilitate rigorous comparison, we release code, exact configs, and per-seed logs via an anonymous link.

Primary Area: learning on graphs and other geometries & topologies

Submission Number: 2672