CoarseFuse: Graph-Coarsening-Based Multi-Atlas Functional Connectivity Fusion for Autism Spectrum Disorder Diagnosis

Autism spectrum disorder (ASD) affects $\sim$1-2% of the population, yet reliable imaging biomarkers remain elusive. Resting-state fMRI (rs-fMRI) enables noninvasive mapping of large-scale connectivity, but single-atlas analyses miss multi-scale effects and many fusion methods trade interpretability for complexity. We present CoarseFuse, a subject-specific, graph-coarsening multi-atlas fusion framework that (i) builds a unified supra-graph from multiple parcellations with space+function cross-atlas affinities, (ii) performs a closed-form, correlation-informed Laplacian refinement with row-sum/PSD projection, and (iii) applies feature-aware local-variation coarsening (LVN) to obtain low-dimensional pseudo-atlases that retain ROI-level interpretability. On ABIDE I, CoarseFuse yields a balanced accuracy (BA) of 82.1% and F1 of 82.0% under stratified 5- fold cross-validation (multiple backbones), outperforming early/late fusion baselines; LVN reduces dimensionality by $\sim$73% (450$\rightarrow$120 nodes). A leave-one-site-out (17-site) evaluation demonstrates robustness to scanner/protocol variation ( macro BA $79.2\%\pm 4.1$; macro F1 $80.1\%\pm 3.9$). Ablations show that explicit cross-atlas edges improve BA by $\sim$1.4-1.5 points and closed-form refinement adds 0.6-0.9 points while improving spectral conditioning. The learned super-nodes align with canonical resting-state networks (e.g., default mode, salience), supporting biological interpretability. To our knowledge, this is the first closed-form Laplacian update tailored for multi-atlas rs-fMRI fusion. CoarseFuse advances rs-fMRI-based ASD diagnosis by combining accuracy, scalability, and transparent network-level insights.