Scholarlink

<p>Abdominal trauma is the most common type of traumatic injury, which results in damage to internal organs and internal bleeding. Abdominal trauma poses significant diagnostic challenges due to its diverse manifestations and potential hidden injuries. This study presents a comprehensive analysis of abdominal trauma classification using 3D CT scans, leveraging advanced deep learning techniques. In this study, we introduce an innovative Dual Attention Transformer architecture, which combines window-attention and multi-head self-attention techniques. This model is intended to properly represent the global as well as local dependencies found in volumetric data. The suggested model is assessed on the RSNA Abdominal Trauma dataset, to identify severe harm in internal abdominal organs such as the liver, kidneys, spleen, and bowel. It is found to be achieving superior performance in contrast to cutting-edge models such as DenseNet, EfficientNet, ResNet, and ViT. Specifically, the proposed model achieves an accuracy of 0.6904, precision of 0.5524, recall of 0.9539, F1-score of 0.7069, and an AUC score of 0.7305, demonstrating superior performance compared to other state-of-the-art models. Our findings show how well the proposed architecture works correctly for identifying abdominal trauma across multiple labels, including Healthy, Injury, Low, and High.</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

Multi-head Self-Attention and Window-Attention Transformer Model for Multi-label Abdominal Trauma Classification Using 3D CT Scans

  • 期刊论文

  • 发表日期：

    2026年3月24日

查看全文
  • Srilatha Chebrolu,
  • Tejo Manasa Vuddagiri,
  • Bhanu Murali Mokkapati,
  • Satvika Dudhyala

摘要

Abdominal trauma is the most common type of traumatic injury, which results in damage to internal organs and internal bleeding. Abdominal trauma poses significant diagnostic challenges due to its diverse manifestations and potential hidden injuries. This study presents a comprehensive analysis of abdominal trauma classification using 3D CT scans, leveraging advanced deep learning techniques. In this study, we introduce an innovative Dual Attention Transformer architecture, which combines window-attention and multi-head self-attention techniques. This model is intended to properly represent the global as well as local dependencies found in volumetric data. The suggested model is assessed on the RSNA Abdominal Trauma dataset, to identify severe harm in internal abdominal organs such as the liver, kidneys, spleen, and bowel. It is found to be achieving superior performance in contrast to cutting-edge models such as DenseNet, EfficientNet, ResNet, and ViT. Specifically, the proposed model achieves an accuracy of 0.6904, precision of 0.5524, recall of 0.9539, F1-score of 0.7069, and an AUC score of 0.7305, demonstrating superior performance compared to other state-of-the-art models. Our findings show how well the proposed architecture works correctly for identifying abdominal trauma across multiple labels, including Healthy, Injury, Low, and High.