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

Area of Science:

• Neuroimaging
• Computational Neuroscience
• Biomedical Engineering

Background:

• Autism spectrum disorder (ASD) affects 1-2% of the population, but reliable imaging biomarkers are lacking.
• Resting-state fMRI (rs-fMRI) maps brain connectivity, but single-atlas analyses have limitations.
• Existing fusion methods often sacrifice interpretability for complexity.

Purpose of the Study:

• To develop a subject-specific, graph-coarsening multi-atlas fusion framework called CoarseFuse.
• To improve the accuracy and interpretability of rs-fMRI analysis for ASD diagnosis.
• To create low-dimensional pseudo-atlases that retain ROI-level interpretability.

Main Methods:

• CoarseFuse builds a unified supra-graph from multiple parcellations using cross-atlas affinities.
• It employs closed-form, correlation-informed Laplacian refinement with projection.
• Feature-aware local-variation coarsening (LVN) generates interpretable, low-dimensional pseudo-atlases.

Main Results:

• CoarseFuse achieved 82.1% balanced accuracy and 82.0% F1 score on the ABIDE I dataset, outperforming baseline fusion methods.
• LVN reduced dimensionality by approximately 73%, from 450 to 120 nodes.
• A leave-one-site-out evaluation confirmed robustness to scanner and protocol variations, with macro BA of 79.2% and macro F1 of 80.1%.

Conclusions:

• CoarseFuse offers an accurate, scalable, and interpretable approach for rs-fMRI-based ASD diagnosis.
• The framework's learned super-nodes align with canonical resting-state networks, supporting biological interpretability.
• This study presents the first closed-form Laplacian update tailored for multi-atlas rs-fMRI fusion, advancing diagnostic capabilities.