Scholarlink

<p>Tongue images play an essential role in Traditional Chinese Medicine (TCM) diagnosis, and accurate tongue segmentation is vital for intelligent tongue analysis. With the growing application of artificial intelligence in tongue diagnosis, robust segmentation is increasingly important for reliable feature extraction and automated syndrome assessment. However, complex backgrounds, large variations in appearance, and small tongue proportions make segmentation highly challenging. This study enhances the U-Net framework by introducing an EC module featuring a residual structure, which aids in stabilising deep feature learning and mitigating the issue of gradient vanishing. We further introduce a Rectangular Self-Calibration Module (RCM) to strengthen foreground extraction and employ an edge-weighted composite loss to improve boundary precision. Experiments across diverse tongue types show that our model achieves more accurate and robust performance than other state-of-the-art methods. Overall, the proposed approach improves IoU by 3.2% over standard U-Net and provides a stronger foundation for practical AI-driven tongue diagnosis systems.</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

TR_Unet: a residual-enhanced U-Net for robust tongue image segmentation in complex conditions

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年3月6日

查看全文
  • Dongsheng Ji,
  • Penghao Chao,
  • Wenhao Fan

摘要

Tongue images play an essential role in Traditional Chinese Medicine (TCM) diagnosis, and accurate tongue segmentation is vital for intelligent tongue analysis. With the growing application of artificial intelligence in tongue diagnosis, robust segmentation is increasingly important for reliable feature extraction and automated syndrome assessment. However, complex backgrounds, large variations in appearance, and small tongue proportions make segmentation highly challenging. This study enhances the U-Net framework by introducing an EC module featuring a residual structure, which aids in stabilising deep feature learning and mitigating the issue of gradient vanishing. We further introduce a Rectangular Self-Calibration Module (RCM) to strengthen foreground extraction and employ an edge-weighted composite loss to improve boundary precision. Experiments across diverse tongue types show that our model achieves more accurate and robust performance than other state-of-the-art methods. Overall, the proposed approach improves IoU by 3.2% over standard U-Net and provides a stronger foundation for practical AI-driven tongue diagnosis systems.