Scholarlink

<p>Chronic diabetic foot and lower limb complications are a major global health concern. India alone has about 77&#xa0;million cases, second only to China with over 116.4&#xa0;million. Infrared thermography provides a useful, non-invasive screening method for differentiating between diabetic and non-diabetic foot conditions by detecting subtle temperature variations that indicate elevated risk of ulceration or amputation. In this study, plantar thermogram images were segmented using an active contour algorithm, and a deep convolutional neural network (CNN) was developed for early diabetic foot detection. A lightweight, custom-designed CNN with three convolutional layers was compared against pretrained models (AlexNet, GoogLeNet, ResNet-18, EfficientNet) under identical preprocessing and evaluation protocols. On the stratified 80/20 split, the proposed CNN achieved 96.84% accuracy, 93.75% sensitivity, 98.41% specificity, 96.77% precision, and 95.23% F1-score. Robustness was confirmed through 5-fold cross-validation, yielding mean values of 95.99% accuracy, 91.88% sensitivity, 98.09% specificity, 96.08% precision, and 93.92% F1-score. For enhanced classification, deep features extracted from CNNs were used with Support Vector Machine (SVM) and K-Nearest Neighbors (KNN) classifiers. The combination of the proposed CNN and SVM achieved the best performance with 97.89% accuracy, 100% sensitivity, 93.75% specificity, 96.92% precision, and 98.44% F1-score. Using plantar thermogram images, this study provides a thorough assessment of cutting-edge deep learning models for precise and timely diabetic foot detection.</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

Deep convolutional neural network for early detection of diabetic foot in plantar thermogram images: a comparative analysis of machine learning approaches

  • 期刊论文

  • 发表日期：

    2026年2月25日

查看全文
  • Saminathan Jayapal,
  • Sasikala Mohan,
  • Nandu Bhavani M,
  • Narayanamurthy V.B.,
  • Rajesh K,
  • Arvind R

摘要

Chronic diabetic foot and lower limb complications are a major global health concern. India alone has about 77 million cases, second only to China with over 116.4 million. Infrared thermography provides a useful, non-invasive screening method for differentiating between diabetic and non-diabetic foot conditions by detecting subtle temperature variations that indicate elevated risk of ulceration or amputation. In this study, plantar thermogram images were segmented using an active contour algorithm, and a deep convolutional neural network (CNN) was developed for early diabetic foot detection. A lightweight, custom-designed CNN with three convolutional layers was compared against pretrained models (AlexNet, GoogLeNet, ResNet-18, EfficientNet) under identical preprocessing and evaluation protocols. On the stratified 80/20 split, the proposed CNN achieved 96.84% accuracy, 93.75% sensitivity, 98.41% specificity, 96.77% precision, and 95.23% F1-score. Robustness was confirmed through 5-fold cross-validation, yielding mean values of 95.99% accuracy, 91.88% sensitivity, 98.09% specificity, 96.08% precision, and 93.92% F1-score. For enhanced classification, deep features extracted from CNNs were used with Support Vector Machine (SVM) and K-Nearest Neighbors (KNN) classifiers. The combination of the proposed CNN and SVM achieved the best performance with 97.89% accuracy, 100% sensitivity, 93.75% specificity, 96.92% precision, and 98.44% F1-score. Using plantar thermogram images, this study provides a thorough assessment of cutting-edge deep learning models for precise and timely diabetic foot detection.