Diabetic foot is a severe complication of diabetes, which can lead to major health issues such as ulceration and amputation if left untreated. While deep learning offers promising avenues for diagnosis, existing studies are often constrained by small, homogeneous datasets, high computational costs, and a tendency to overfit to single data sources. To address these limitations, this study develops a robust classification pipeline that integrates multiple plantar thermography datasets to improve model generalizability and accessibility. The pipeline focuses on systematic data integration, preprocessing, and validation, training a suite of lightweight convolutional neural networks (CNNs) with transfer learning on standardized and combined thermograms. Model explainability was verified using Grad-CAM visualizations, which confirmed that predictions were based on physiologically relevant plantar regions rather than background artifacts. The best-performing architecture within this pipeline, GhostNet₁00, achieved new state-of-the-art performance on multi-dataset evaluation with overall mean values of 96. 0% accuracy, 97. 1% precision, 97. 9% recall, 97. 5% F1-score, and 94. 3% specificity. By unifying diverse datasets and deploying efficient CNNs within a reproducible workflow, this work establishes a scalable, cost-effective pipeline for early diabetic foot screening, demonstrating strong potential for practical deployment in low-resource healthcare environments. • Thermal colormap standardization enables multi-dataset integration. • Multi-dataset integration improves diabetic foot detection generalization. • GhostNet₁00 transfer learning achieves SOTA scores on combined datasets. • Grad-CAM visualizations improve explainability and clinical reliability. • Mobile deployment feasibility supports low-resource healthcare settings.
Tahmid et al. (Tue,) studied this question.