Scholarlink

<p>Diabetic foot disease is a major complication of diabetes and a leading cause of preventable amputations. This study investigates multimodal imaging for early diabetic foot risk assessment using artificial intelligence. Using the STANDUP dataset of 415 paired RGB and infrared (IR) plantar images from healthy individuals and patients with type II diabetes, deep learning models were trained on RGB images, IR images, and their combinations. Hybrid approaches combining CNN predictions with interpretable thermal features were evaluated. IR-based models consistently outperformed RGB-only approaches: The best-performing model achieved 99.5% balanced accuracy, 99.7% precision, 99.7% recall and 99.7% F1-score in binary classification. Hybrid CNN-statistical-feature models improved both binary classification and multi-class risk stratification, with the top hybrid configuration reaching 100% balanced accuracy, precision, recall, and F1-score in the binary task. Models trained solely on binary foot-shape masks achieved balanced accuracy up to 93.4%, precision of 95.1%, recall of 94.7%, and F1-score of 94.7%, highlighting the diagnostic relevance of foot morphology. By relying on fully automated image segmentation and widely available RGB and infrared imaging, the framework is scalable and low-cost, suggesting its potential use in population-level screening after external validation.</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

Machine learning approaches for diabetic plantar foot classification from RGB and thermal images

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年7月10日

查看全文
  • Lorenzo Muscillo,
  • Corin F. Otesteanu,
  • Carlo Menon

摘要

Diabetic foot disease is a major complication of diabetes and a leading cause of preventable amputations. This study investigates multimodal imaging for early diabetic foot risk assessment using artificial intelligence. Using the STANDUP dataset of 415 paired RGB and infrared (IR) plantar images from healthy individuals and patients with type II diabetes, deep learning models were trained on RGB images, IR images, and their combinations. Hybrid approaches combining CNN predictions with interpretable thermal features were evaluated. IR-based models consistently outperformed RGB-only approaches: The best-performing model achieved 99.5% balanced accuracy, 99.7% precision, 99.7% recall and 99.7% F1-score in binary classification. Hybrid CNN-statistical-feature models improved both binary classification and multi-class risk stratification, with the top hybrid configuration reaching 100% balanced accuracy, precision, recall, and F1-score in the binary task. Models trained solely on binary foot-shape masks achieved balanced accuracy up to 93.4%, precision of 95.1%, recall of 94.7%, and F1-score of 94.7%, highlighting the diagnostic relevance of foot morphology. By relying on fully automated image segmentation and widely available RGB and infrared imaging, the framework is scalable and low-cost, suggesting its potential use in population-level screening after external validation.