Enhancing Diabetic Foot Ulcer Classification Through Fine-Tuned Multilevel CNN

Diabetic foot ulcers (DFUs) are especially problematic with their high rates of recurrence. Recognizing DFUs with infection and ischemia is vital for appropriate treatment and prevention of amputations. The proposed approach combines two convolutional neural networks (CNNs), fine-tuned through multilevel refinement, to classify DFU images into four categories: nonexistent, infection, ischemia, and both (infection/ischemia). Transfer learning was used and the adopted multilevel design includes fully connected layers with varying numbers of neurons and batch normalization. Data augmentation techniques were applied to a dataset comprising 8,242 images to deal with the class imbalance problem, mitigate the overfitting possibility, and enhance the approach’s performance. Based on a five-fold cross-validation, the proposed approach obtained 95.91%, 93.28%, and 95.1% of accuracy, Kappa, and F1-score metrics, respectively, confirming its superiority compared to the current solutions for the same number of image classes. Furthermore, a cross-dataset validation method was employed to evaluate the model’s generalization ability. Employing this methodology, the Kaggle DFU dataset obtained accuracy, Kappa, and F1-score values of 79.79%, 59.80%, and 79.02%. Therefore, the proposed approach appears promising for the precise and timely diagnosis of DFU images.