Scholarlink

<p>Tongue diagnosis is a crucial component of the Four Diagnostic Methods in Traditional Chinese Medicine (TCM), which include observing, listening and smelling, inquiring, and palpation. Tongue image segmentation holds great significance in advancing the intelligentization of tongue diagnosis research. This paper introduces an improved model called Parallel Attention and Progressive Upsampling for Tongue Segmentation (PAPU_TonSeg), based on the Segformer architecture, to address the issues of inaccurate and blurred tongue edge segmentation in tongue semantic segmentation. The model incorporates three key enhancements: (1) the adoption of a Self-Attention Parallel Network that integrates the self-attention mechanism and residual modules to achieve simultaneous extraction of local and global features; (2) the integration of the Efficient Channel Attention(ECA) mechanism into the Mix-FFN component to enhance feature extraction efficiency; and (3) the utilization of Multi-dimensional Feature Progressive Upsampling to mitigate precision loss during the upsampling process. Evaluation results on the BioHit public dataset demonstrate that, compared to the original Segformer, PAPU_TonSeg achieves improvements of 2.42% in Mean Pixel Accuracy (MPA), 0.78% in Mean Intersection over Union (MIoU), and 2.02% in the Dice coefficient, while boasting a lower parameter count and computational complexity. On another dataset, PAPU_TonSeg outperforms Segformer with an MPA increase of 0.64%, an MIoU increase of 0.33%, and a Dice coefficient increase of 0.4%. The improved model not only has fewer parameters but also exhibits a notably lower computational complexity compared to classical models. The PAPU_TonSeg model accurately segments tongue body details, such as tooth marks, and distributes attention more evenly, capturing both global and local features. These findings position PAPU_TonSeg as a valuable tool for clinical diagnosis and research in TCM tongue diagnosis.</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

An improved lightweight tongue segmentation model with self-attention parallel network and progressive upsampling

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年7月29日

查看全文
  • Xuan Wang,
  • Yifang Cao,
  • Yijia Chen,
  • Huixia Li,
  • Aiqing Han,
  • Yan Tang

摘要

Tongue diagnosis is a crucial component of the Four Diagnostic Methods in Traditional Chinese Medicine (TCM), which include observing, listening and smelling, inquiring, and palpation. Tongue image segmentation holds great significance in advancing the intelligentization of tongue diagnosis research. This paper introduces an improved model called Parallel Attention and Progressive Upsampling for Tongue Segmentation (PAPU_TonSeg), based on the Segformer architecture, to address the issues of inaccurate and blurred tongue edge segmentation in tongue semantic segmentation. The model incorporates three key enhancements: (1) the adoption of a Self-Attention Parallel Network that integrates the self-attention mechanism and residual modules to achieve simultaneous extraction of local and global features; (2) the integration of the Efficient Channel Attention(ECA) mechanism into the Mix-FFN component to enhance feature extraction efficiency; and (3) the utilization of Multi-dimensional Feature Progressive Upsampling to mitigate precision loss during the upsampling process. Evaluation results on the BioHit public dataset demonstrate that, compared to the original Segformer, PAPU_TonSeg achieves improvements of 2.42% in Mean Pixel Accuracy (MPA), 0.78% in Mean Intersection over Union (MIoU), and 2.02% in the Dice coefficient, while boasting a lower parameter count and computational complexity. On another dataset, PAPU_TonSeg outperforms Segformer with an MPA increase of 0.64%, an MIoU increase of 0.33%, and a Dice coefficient increase of 0.4%. The improved model not only has fewer parameters but also exhibits a notably lower computational complexity compared to classical models. The PAPU_TonSeg model accurately segments tongue body details, such as tooth marks, and distributes attention more evenly, capturing both global and local features. These findings position PAPU_TonSeg as a valuable tool for clinical diagnosis and research in TCM tongue diagnosis.