Scholarlink

Chronic wound assessment requires accurate segmentation and interpretable analysis for effective treatment planning. Although deep learning models show promise in automated wound segmentation, clinical adoption is hampered by limited interpretability. In this paper, we present a framework that integrates advanced wound segmentation with conversational interfaces powered by a large language model (LLM). Our EfficientNet-Attention U-Net trained on 2,760 wound images achieves state-of-the-art performance with an accuracy of 99.76%, Dice score of 0.9065, and intersection over Union (IoU) of 0.8324. Beyond segmentation, we extract morphological features, color distributions, texture properties, and edge characteristics. We integrate GPT 2.0 through an interface that translates quantitative measurements into clinical narratives, allowing healthcare providers to naturally interact with AI assessments, by We integrate GPT 2.0 to enable a conversational AI system that provides healthcare providers with clear, clinically meaningful explanations of extracted wound properties. The LLM-generated explanations were evaluated by healthcare professionals and received high scores across multiple dimensions. Interpretability and clarity received high scores, with a median of 4.33 (IQR 4.0–5.0), indicating that most respondents found the explanations easy to understand and well-presented and 85.2% of participants agreed this assessment. Consistency, reflecting the logical coherence and reliability of information across explanations, achieved a median rating of 4.00 (IQR 3.7–4.7), showing moderate variability and 72.8% agreement among evaluators. Usefulness, capturing how informative and actionable the explanations were also received a median of 4.00 (IQR 4.0–4.7), with 77.8% agreement, indicating generally favorable evaluations. Finally, the absence of unsupported or hallucinated information was rated positively, with a median of 4.00 (IQR 4.0–5.0) and 91.4% agreement, indicating that most participants found the explanations reliable and factually accurate. The results demonstrate superior segmentation accuracy alongside clinically meaningful interpretations, addressing the need for explainable AI in healthcare. Our system offers a practical solution for remote and clinical wound care settings.

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 Based Wound Segmentation and Characterization with LLM-Assisted Clinical Interpretability

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Fisha Mehabaw Alemayoh,
  • Hailemicael Lulseged Yimer,
  • Xiaohong Yuan,
  • Letu Qingge

摘要

Chronic wound assessment requires accurate segmentation and interpretable analysis for effective treatment planning. Although deep learning models show promise in automated wound segmentation, clinical adoption is hampered by limited interpretability. In this paper, we present a framework that integrates advanced wound segmentation with conversational interfaces powered by a large language model (LLM). Our EfficientNet-Attention U-Net trained on 2,760 wound images achieves state-of-the-art performance with an accuracy of 99.76%, Dice score of 0.9065, and intersection over Union (IoU) of 0.8324. Beyond segmentation, we extract morphological features, color distributions, texture properties, and edge characteristics. We integrate GPT 2.0 through an interface that translates quantitative measurements into clinical narratives, allowing healthcare providers to naturally interact with AI assessments, by We integrate GPT 2.0 to enable a conversational AI system that provides healthcare providers with clear, clinically meaningful explanations of extracted wound properties. The LLM-generated explanations were evaluated by healthcare professionals and received high scores across multiple dimensions. Interpretability and clarity received high scores, with a median of 4.33 (IQR 4.0–5.0), indicating that most respondents found the explanations easy to understand and well-presented and 85.2% of participants agreed this assessment. Consistency, reflecting the logical coherence and reliability of information across explanations, achieved a median rating of 4.00 (IQR 3.7–4.7), showing moderate variability and 72.8% agreement among evaluators. Usefulness, capturing how informative and actionable the explanations were also received a median of 4.00 (IQR 4.0–4.7), with 77.8% agreement, indicating generally favorable evaluations. Finally, the absence of unsupported or hallucinated information was rated positively, with a median of 4.00 (IQR 4.0–5.0) and 91.4% agreement, indicating that most participants found the explanations reliable and factually accurate. The results demonstrate superior segmentation accuracy alongside clinically meaningful interpretations, addressing the need for explainable AI in healthcare. Our system offers a practical solution for remote and clinical wound care settings.