Scholarlink

<p>Coronary artery disease (CAD) is a leading cause of mortality worldwide, with its insidious nature and high fatality rate emphasizing the critical importance of early screening. However, current facial video-based CAD screening methods struggle with feature extraction, making it hard to capture subtle long-term facial variations. To address this issue, this paper proposes CADMamba, a deep neural network model based on the Mamba module. CADMamba designed to assess CAD risk non-invasively by analyzing easily accessible facial videos from daily life. To enhance feature extraction and capture long-term temporal dependencies, CADMamba incorporates a differential frame feature fusion method based on self-attention (DFF-AFF) and the CADMamba Encoder. The model was evaluated using 2,250 video samples (374 subjects, of whom 54% had CAD), demonstrating a strong correlation between CADMamba’s predictions and actual outcomes. The recall and precision rates were 84.70% and 80.37%, respectively, representing a notable improvement compared to baseline video models. This model shows promise in supporting early CAD screening efforts, especially in areas with scarce medical resources.</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

Cadmamba: a differential feature fusion-based neural network for coronary artery disease screening from facial videos

  • 期刊论文

  • 发表日期：

    2025年4月2日

查看全文
  • Haohan Ou,
  • Baoping Jia,
  • Yishan Hu,
  • Meiling Cai,
  • Yan Qiang,
  • Qi Chen,
  • Bo Pei,
  • Juanjuan Zhao

摘要

Coronary artery disease (CAD) is a leading cause of mortality worldwide, with its insidious nature and high fatality rate emphasizing the critical importance of early screening. However, current facial video-based CAD screening methods struggle with feature extraction, making it hard to capture subtle long-term facial variations. To address this issue, this paper proposes CADMamba, a deep neural network model based on the Mamba module. CADMamba designed to assess CAD risk non-invasively by analyzing easily accessible facial videos from daily life. To enhance feature extraction and capture long-term temporal dependencies, CADMamba incorporates a differential frame feature fusion method based on self-attention (DFF-AFF) and the CADMamba Encoder. The model was evaluated using 2,250 video samples (374 subjects, of whom 54% had CAD), demonstrating a strong correlation between CADMamba’s predictions and actual outcomes. The recall and precision rates were 84.70% and 80.37%, respectively, representing a notable improvement compared to baseline video models. This model shows promise in supporting early CAD screening efforts, especially in areas with scarce medical resources.