Cross-attention-guided subject-adaptive graph learning for multimodal autism classification: integrating structural...

Area of Science:

• Neuroscience
• Medical Imaging
• Machine Learning

Background:

• Autism spectrum disorder (ASD) presents as a complex neurodevelopmental condition characterized by atypical brain structure and function.
• Diagnosing ASD is challenging due to individual variability and inconsistencies across different data collection sites.

Purpose of the Study:

• To develop an advanced computational model for diagnosing ASD by integrating multimodal neuroimaging data.
• To enhance the accuracy and generalizability of ASD diagnostic markers, addressing data variability issues.

Main Methods:

• Proposed a cross-attention-guided subject-adaptive graph network (CAS-GNN) model.
• Integrated structural MRI and resting-state functional connectivity data for comprehensive brain network analysis.
• Employed a site-invariant learning strategy to improve cross-site generalization.

Main Results:

• The CAS-GNN model achieved high diagnostic accuracy on the ABIDE-I dataset, reaching 79.25% ± 4.71% on independent test data.
• Demonstrated superior performance compared to existing machine learning baselines.
• Identified specific brain regions and connections associated with ASD, highlighting right-hemisphere dominance.

Conclusions:

• The CAS-GNN framework offers a robust and interpretable method for ASD diagnosis.
• The model provides valuable neurobiological insights into ASD pathophysiology.
• Accelerates the discovery and development of reliable biomarkers for ASD.