Scholarlink

<p>Tongue diagnosis is important in Traditional Chinese Medicine (TCM), but routine inspection remains subjective and inconsistent across observers. Real outpatient images add background clutter, illumination variation, and label ambiguity. We propose RTDS, a two-stage framework for noisy clinical data: Stage 1 uses U-Net++ for tongue ROI extraction, and Stage 2 uses a ResNet-34 + Swin hybrid classifier with Focal Loss for imbalance- and ambiguity-aware learning. On 2,100 real outpatient images, RTDS achieves 75.76% Top-1 accuracy for seven-class tongue appearance classification and 91.30% accuracy after mapping to four diagnostic states, outperforming strong baselines under the same protocol. Grad-CAM results show attention concentrated on clinically relevant tongue body/coating regions, supporting practical interpretability.</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

RTDS: A Robust two-stage tongue diagnosis system with U-Net++ segmentation and swin-hybrid classification

  • 期刊论文

  • 发表日期：

    2026年4月13日

查看全文
  • Boyang Zhang

摘要

Tongue diagnosis is important in Traditional Chinese Medicine (TCM), but routine inspection remains subjective and inconsistent across observers. Real outpatient images add background clutter, illumination variation, and label ambiguity. We propose RTDS, a two-stage framework for noisy clinical data: Stage 1 uses U-Net++ for tongue ROI extraction, and Stage 2 uses a ResNet-34 + Swin hybrid classifier with Focal Loss for imbalance- and ambiguity-aware learning. On 2,100 real outpatient images, RTDS achieves 75.76% Top-1 accuracy for seven-class tongue appearance classification and 91.30% accuracy after mapping to four diagnostic states, outperforming strong baselines under the same protocol. Grad-CAM results show attention concentrated on clinically relevant tongue body/coating regions, supporting practical interpretability.