Scholarlink

<p>Diabetes mellitus is a significant burden in over 537 million individuals globally, and neuropathy and vascular deterioration are common to occur in the lower extremities. Using a plantar thermographic imaging technique, this article presents OZAIANet, a novel deep learning model designed especially for non-invasive diabetes diagnosis. The model utilizes an improved combination of DenseNet121 and EfficientNetV2B0 along with SHAP and LIME explainability techniques to impart clinical interpretability. OZAIANet was trained on a benchmark dataset of 167 individuals (122 diabetic, 45 non-diabetic) and performed significantly better than baseline models in terms of classification accuracy of 94.34%, precision of 96.72%, recall of 93.65%, and F1-score of 95.11%. A thorough comparative analysis validated the excellence of the pairwise fusion approach compared to individual CNNs. The employment of dual explainable AI mechanisms also highlighted the model’s capability for identifying pathophysiologically significant areas in diabetic foot radiographs. The end-to-end, interpretable, and high-performance system demonstrates strong potential for early diabetes screening and future translation into clinical settings, subject to large-scale 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

OZAIANet a fusion based explainable deep learning framework for thermal image based diabetes classification

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年7月28日

查看全文
  • Saksham Anand,
  • Suganya R,
  • M. Vimudha

摘要

Diabetes mellitus is a significant burden in over 537 million individuals globally, and neuropathy and vascular deterioration are common to occur in the lower extremities. Using a plantar thermographic imaging technique, this article presents OZAIANet, a novel deep learning model designed especially for non-invasive diabetes diagnosis. The model utilizes an improved combination of DenseNet121 and EfficientNetV2B0 along with SHAP and LIME explainability techniques to impart clinical interpretability. OZAIANet was trained on a benchmark dataset of 167 individuals (122 diabetic, 45 non-diabetic) and performed significantly better than baseline models in terms of classification accuracy of 94.34%, precision of 96.72%, recall of 93.65%, and F1-score of 95.11%. A thorough comparative analysis validated the excellence of the pairwise fusion approach compared to individual CNNs. The employment of dual explainable AI mechanisms also highlighted the model’s capability for identifying pathophysiologically significant areas in diabetic foot radiographs. The end-to-end, interpretable, and high-performance system demonstrates strong potential for early diabetes screening and future translation into clinical settings, subject to large-scale validation.