A functional system-informed graph neural network framework to quantify interpretable brain dysfunction in ASD.

Area of Science:

• Neuroscience
• Artificial Intelligence
• Medical Imaging

Background:

• Autism Spectrum Disorder (ASD) diagnosis relies on subjective assessments, lacking definitive pathological understanding.
• Brain connectivity and functional systems are crucial for understanding neurodevelopmental disorders.
• Current diagnostic methods for ASD need improvement in accuracy and objectivity.

Purpose of the Study:

• To develop an advanced AI framework for diagnosing ASD using functional magnetic resonance imaging (fMRI).
• To leverage brain connectome topology and function for enhanced ASD detection.
• To uncover novel insights into ASD-related brain dysfunctions.

Main Methods:

• Proposed a functional system-informed graph neural network (FS-GNN) framework.
• Integrated learnable positional encoding for brain regions of interest (ROIs).
• Incorporated large-scale brain systems as prior knowledge and used system-driven regularization for ROI weighting.

Main Results:

• FS-GNN achieved 75.02% accuracy, 73.22% precision, and 71.64% recall in ASD diagnosis on the ABIDE dataset.
• Outperformed existing machine learning and GNN approaches.
• Identified brain dysfunctions at ROI and system levels consistent with known ASD biomarkers.

Conclusions:

• FS-GNN demonstrates efficacy in diagnosing ASD by utilizing brain connectome information.
• The framework provides interpretable and trustworthy neural patterns for precise ASD diagnosis.
• This approach offers a promising direction for objective ASD assessment and understanding.