DFU-XAI: Explainable DFU Classification with Efficientnet-B3, A Grad-CAM-Enhanced Deep Learning Framework

Diabetic foot ulcer (DFU) represents a critical diabetes-related health threat that results in both extensive lower-limb amputation cases worldwide and diabetes complications. Deep learning-based diagnostic screening systems deliver impressive diagnostic results but remain clinically questionable because they lack interpretability. The research introduces an interpretable DFU detection system that combines fine-tuned EfficientNet-B3 base model with Grad-CAM for binary classification tasks. The authors conducted tests on the publicly accessible Kaggle DFU dataset and developed the system through training on image parts and conducted testing on an independent dataset. The developed model produced perfect test accuracy (97 %) paired with identical scores for precision, recall, specificity and F1 (all 97%) and an AUC score of 0.99. The research demonstrates the proposed model's superior performance over MobileNetV2 and ResNet50 base models. The model's transparency increases due to Grad-CAM visualizations which reveal medically significant ulcer locations. The DFU-XAI screening and triage framework presents users with both reliable performance and clear system interpretability.