Scholarlink

<p>Autism Spectrum Disorder (ASD) is a complex neurodevelopmental condition characterized by persistent social communication difficulties and restricted, repetitive behaviors. Despite growing interest in machine learning and deep learning approaches based on functional and structural magnetic resonance imaging, reliable ASD diagnosis remains challenging. We propose a Multimodal Graph Convolutional Network with an Attention Mechanism (MGCN-AM) that integrates complementary information from fMRI, sMRI, and phenotypic data. The MGCN-AM constructs three group-level graphs and employs multiple fusion attention convolution layers to learn cross-modality interactions and extract salient features. A multilayer perceptron (MLP) is incorporated to perform the final classification. The model was evaluated on the ABIDE I dataset comprising 871 subjects with both fMRI and sMRI modalities. Performance was assessed using 10-fold cross-validation while maintaining site proportions across 17 sites. Under this setting, MGCN-AM achieved an average accuracy of 88.4%, sensitivity of 87.4%, and specificity of 89.8%. The proposed method outperforms existing state-of-the-art models and demonstrates clear advantages over single-modality approaches, highlighting the effectiveness of graph-based multimodal fusion for ASD diagnosis.</p>

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

A multi-graph convolution network with attention mechanism based on multi-modal data for ASD diagnosis

  • 期刊论文

  • 发表日期：

    2026年4月2日

查看全文
  • Lizhen Shao,
  • Jian Xiao,
  • Kexin Shi,
  • Dongmei Fu

摘要

Autism Spectrum Disorder (ASD) is a complex neurodevelopmental condition characterized by persistent social communication difficulties and restricted, repetitive behaviors. Despite growing interest in machine learning and deep learning approaches based on functional and structural magnetic resonance imaging, reliable ASD diagnosis remains challenging. We propose a Multimodal Graph Convolutional Network with an Attention Mechanism (MGCN-AM) that integrates complementary information from fMRI, sMRI, and phenotypic data. The MGCN-AM constructs three group-level graphs and employs multiple fusion attention convolution layers to learn cross-modality interactions and extract salient features. A multilayer perceptron (MLP) is incorporated to perform the final classification. The model was evaluated on the ABIDE I dataset comprising 871 subjects with both fMRI and sMRI modalities. Performance was assessed using 10-fold cross-validation while maintaining site proportions across 17 sites. Under this setting, MGCN-AM achieved an average accuracy of 88.4%, sensitivity of 87.4%, and specificity of 89.8%. The proposed method outperforms existing state-of-the-art models and demonstrates clear advantages over single-modality approaches, highlighting the effectiveness of graph-based multimodal fusion for ASD diagnosis.