Scholarlink

<p>Early diagnosis of autism spectrum disorder (ASD) is crucial for effective treatment planning and improved patient outcomes. Functional magnetic resonance imaging (fMRI) serves as a valuable tool for identifying atypical brain connectivity patterns associated with ASD. However, the high dimensionality and complexity of fMRI data make accurate classification and severity assessment challenging. To address this, the present study proposes a two-stage deep learning framework based on fMRI brain scans to automate the diagnosis of ASD and its severity levels. In the first stage, a Swin3D Transformer model is employed to perform binary classification of ASD and non-ASD cases. For ASD-positive cases, the second stage utilizes a Hybrid Graph Attention Network (GAT) integrated with a Long Short-Term Memory (LSTM) model to classify severity levels into mild, moderate, and severe categories. The proposed framework achieves an accuracy of 93.8% in binary ASD detection and 92.3% in severity-level classification, indicating the effectiveness of the model in capturing both spatial and temporal dependencies in brain activity. Thus, the approach provides a reliable and interpretable tool for clinicians working in the challenging area of early and precise diagnosis of ASD, to aid the implementation of more personalized and timely therapeutic interventions.</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

Dual-Stage Deep Learning Framework for Autism Spectrum Disorder Detection and Severity Classification Using Swin3D Transformer and GAT-LSTM Networks

  • 期刊论文

  • 发表日期：

    2025年12月24日

查看全文
  • A. Padmapriya,
  • D. Suganya,
  • R. Kalpana

摘要

Early diagnosis of autism spectrum disorder (ASD) is crucial for effective treatment planning and improved patient outcomes. Functional magnetic resonance imaging (fMRI) serves as a valuable tool for identifying atypical brain connectivity patterns associated with ASD. However, the high dimensionality and complexity of fMRI data make accurate classification and severity assessment challenging. To address this, the present study proposes a two-stage deep learning framework based on fMRI brain scans to automate the diagnosis of ASD and its severity levels. In the first stage, a Swin3D Transformer model is employed to perform binary classification of ASD and non-ASD cases. For ASD-positive cases, the second stage utilizes a Hybrid Graph Attention Network (GAT) integrated with a Long Short-Term Memory (LSTM) model to classify severity levels into mild, moderate, and severe categories. The proposed framework achieves an accuracy of 93.8% in binary ASD detection and 92.3% in severity-level classification, indicating the effectiveness of the model in capturing both spatial and temporal dependencies in brain activity. Thus, the approach provides a reliable and interpretable tool for clinicians working in the challenging area of early and precise diagnosis of ASD, to aid the implementation of more personalized and timely therapeutic interventions.