Scholarlink

<p>Accurate evaluation of coronary intermediate lesions (50–70% stenosis) is essential for stent decision-making, yet conventional angiography remains subjective and adjunctive tests like FFR are often invasive or costly. In this retrospective multicenter study of 1298 patients, we developed an attention-enhanced deep learning model using Improved_EfficientNet with a Convolutional Block Attention Module to predict stent necessity directly from coronary angiography images. The model utilized multimodal labels from FFR, IVUS, and OCT as reference standards during training. In internal validation, the model achieved an accuracy of 0.976 and an F1-score of 0.971. External validation across independent institutions demonstrated robust performance with an accuracy of 0.807 and an AUC of 0.897. Grad-CAM visualization confirmed that the model focuses on clinically relevant stenotic regions, showing high alignment with expert interpretations. These results suggest that the proposed model can effectively integrate anatomical and functional information to provide real-time decision support, potentially reducing the need for invasive adjunctive testing and enhancing precision in interventional cardiology.</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

Deep learning predicts stent implantation in borderline coronary lesions from angiography

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年5月22日

查看全文
  • Jingsong Xia,
  • Di Zhao,
  • Yiming Zhang,
  • Leilei Chen,
  • Zhenhua Yang,
  • Dengqing Shi,
  • Chao Liu,
  • Haoyu Meng,
  • Liansheng Wang,
  • Jiabao Liu

摘要

Accurate evaluation of coronary intermediate lesions (50–70% stenosis) is essential for stent decision-making, yet conventional angiography remains subjective and adjunctive tests like FFR are often invasive or costly. In this retrospective multicenter study of 1298 patients, we developed an attention-enhanced deep learning model using Improved_EfficientNet with a Convolutional Block Attention Module to predict stent necessity directly from coronary angiography images. The model utilized multimodal labels from FFR, IVUS, and OCT as reference standards during training. In internal validation, the model achieved an accuracy of 0.976 and an F1-score of 0.971. External validation across independent institutions demonstrated robust performance with an accuracy of 0.807 and an AUC of 0.897. Grad-CAM visualization confirmed that the model focuses on clinically relevant stenotic regions, showing high alignment with expert interpretations. These results suggest that the proposed model can effectively integrate anatomical and functional information to provide real-time decision support, potentially reducing the need for invasive adjunctive testing and enhancing precision in interventional cardiology.