Scholarlink

Autism Spectrum Disorder (ASD) diagnosis faces challenges from subjective behavioral assessments, highlighting need for objective neuroimaging approaches. Current methods face limitations in integrating multi-modal brain connectivity patterns and achieving clinically viable performance. We propose SMART-HetGNN, a framework integrating Selective Multi-Atlas ROI Transformer (SMART) mechanisms with heterogeneous graph neural networks for automated ASD diagnosis. SMART adaptively extracts connectivity features across brain atlases (AAL, CC200, Harvard-Oxford), while dual transformer architecture with cross-attention fusion processes multi-atlas features effectively. A heterogeneous graph neural network models complex interactions between subject features, ROI connectivity patterns, and phenotypic information through gated attention mechanisms. Experiments on ABIDE dataset demonstrate exceptional performance with 96.2% accuracy, 0.991 AUC, and perfect precision (100%), representing substantial improvements over state-of-the-art methods. Cross-site validation reveals strong generalization with minimal performance degradation (average 2.1% drop), significantly outperforming typical neuroimaging studies. Ablation studies validate importance of multi-atlas integration and phenotypic information. The framework offers potential for clinical deployment, providing objective ASD diagnosis with perfect precision while maintaining high sensitivity.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

SMART-HetGNN: A Novel Selective Multi-atlas ROI Transformer Heterogeneous Graph Neural Network for Autism Spectrum Disorder Diagnosis

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Dashu Zhang

摘要

Autism Spectrum Disorder (ASD) diagnosis faces challenges from subjective behavioral assessments, highlighting need for objective neuroimaging approaches. Current methods face limitations in integrating multi-modal brain connectivity patterns and achieving clinically viable performance. We propose SMART-HetGNN, a framework integrating Selective Multi-Atlas ROI Transformer (SMART) mechanisms with heterogeneous graph neural networks for automated ASD diagnosis. SMART adaptively extracts connectivity features across brain atlases (AAL, CC200, Harvard-Oxford), while dual transformer architecture with cross-attention fusion processes multi-atlas features effectively. A heterogeneous graph neural network models complex interactions between subject features, ROI connectivity patterns, and phenotypic information through gated attention mechanisms. Experiments on ABIDE dataset demonstrate exceptional performance with 96.2% accuracy, 0.991 AUC, and perfect precision (100%), representing substantial improvements over state-of-the-art methods. Cross-site validation reveals strong generalization with minimal performance degradation (average 2.1% drop), significantly outperforming typical neuroimaging studies. Ablation studies validate importance of multi-atlas integration and phenotypic information. The framework offers potential for clinical deployment, providing objective ASD diagnosis with perfect precision while maintaining high sensitivity.