Journal of Health and Biomedical Informatics
Volume 12, Issue 4 (3-2026)
jhbmi 2026, 12(4): 315-333 |
Back to browse issues page
Download citation:
BibTeX | RIS | EndNote | Medlars | ProCite | Reference Manager | RefWorks
Send citation to:
BibTeX | RIS | EndNote | Medlars | ProCite | Reference Manager | RefWorks
Send citation to:
Ahmadi Rad Z, Gheyoumi Zadeh H, Fayazi A, Rezaee K. Integration of Image and Tabular Data in a Siamese Network for Diabetic Patient Classification Using Thermography: A Multimodal EfficientNetB3-Based Approach. jhbmi 2026; 12 (4) :315-333
URL: http://jhbmi.ir/article-1-953-en.html
URL: http://jhbmi.ir/article-1-953-en.html
Ph.D. in Biomedical Engineering, Associate Professor, Department of Electrical Engineering, Faculty of Engineering, Vali-e-Asr University of Rafsanjan, Rafsanjan, Iran
Abstract: (870 Views)
Introduction: Diabetic foot is a serious chronic complication of diabetes that can lead to ulcers and limb amputation if not diagnosed early. while non-invasive techniques like thermography can prevent severe complications, traditional methods often fail to identify subtle thermal changes. This study developed an intelligent system for diabetic foot detection using thermographic images and numerical features.
Method: The Plantar Thermogram Database, containing 167 samples, was divided into training (116 samples), validation (25 samples), and test (26 samples) sets. A multimodal Siamese architecture based on EfficientNetB3 processed left and right foot thermograms simultaneously. Outputs were combined using ∣L−R∣ and L⊙R to model thermal asymmetry, then fused with normalized tabular features. Implementation used TensorFlow and Keras, with class weighting to address data imbalance.
Results: The model achieved 96.15% accuracy, 95% sensitivity, 100% specificity, 100% precision, 85.71% NPV, 97.44% F1-score, and 97.5% balanced accuracy on the test set. Learning curves showed rapid convergence with no overfitting and stable performance.
Conclusion: The proposed Siamese architecture, utilizing both image and tabular data, offers a non-invasive, rapid, and accurate approach for screening and monitoring diabetic patients, potentially facilitating early diagnostic clinical systems for diabetic foot detection.
Method: The Plantar Thermogram Database, containing 167 samples, was divided into training (116 samples), validation (25 samples), and test (26 samples) sets. A multimodal Siamese architecture based on EfficientNetB3 processed left and right foot thermograms simultaneously. Outputs were combined using ∣L−R∣ and L⊙R to model thermal asymmetry, then fused with normalized tabular features. Implementation used TensorFlow and Keras, with class weighting to address data imbalance.
Results: The model achieved 96.15% accuracy, 95% sensitivity, 100% specificity, 100% precision, 85.71% NPV, 97.44% F1-score, and 97.5% balanced accuracy on the test set. Learning curves showed rapid convergence with no overfitting and stable performance.
Conclusion: The proposed Siamese architecture, utilizing both image and tabular data, offers a non-invasive, rapid, and accurate approach for screening and monitoring diabetic patients, potentially facilitating early diagnostic clinical systems for diabetic foot detection.
Keywords: Diabetic foot, Thermography, Siamese architecture, Multimodal data, Early detection, EfficientNetB3
Type of Study: Original Article |
Subject:
Artificial Intelligence in Healthcare
Received: 2025/09/18 | Accepted: 2026/02/15
Received: 2025/09/18 | Accepted: 2026/02/15
Send email to the article author
| Rights and permissions | |
 |
This work is licensed under a Creative Commons Attribution-NonCommercial 4.0 International License. |