Publication

Abstract :

Diabetic Foot Ulcer (DFU) is a serious complication of diabetes and affects millions worldwide. It results in amputation if left undiagnosed and untreated. Conventional clinical diagnosis is based on mere visual examination, which is prone to subjectivity and inter-observer variability. This work presents a computer-aided diagnosis system using EfficientNet architectures (B0-B3) for detection and visualization in DFU images. These models have been trained using a curated openly available dataset consisting of 543 normal and 512 ulcerated skin patches with extensive data augmentation, transfer learning, and the proposed custom classification head as techniques to enhance performance for binary classification. Model evaluation includes predictive metrics such as accuracy, precision, recall, and F1score with computational metrics like FLOPs, parameters count, and inference time, giving a complete picture of accuracy and efficiency. Interpretability is enhanced through Grad-CAM visualizations, highlighting clinically relevant areas that influence the model's decisions. Experimental results showed that EfficientNetB3 yielded the highest test accuracy of 99.37 % with near-perfect precision, recall, and F1-score of 0.994. Among lighter variants, EfficientNet-B0 and B2 obtained competitive accuracy (98.74 %) with very low computational costs and were thus suitable for resource-constrained deployment, whereas EfficientNet-B1 obtained 98.11 % accuracy with a balanced performance but moderately higher inference latency. This study also illustrates how EfficientNet-based models can be used to achieve accurate, efficient, and interpretable detection of DFUs.