Pretrained Deep Learning Networks for Diabetic Foot Ulcer (DFU) Classification

Diabetic foot ulcers (DFUs) present significant challenges in diagnosis, necessitating accurate classification of conditions, such as ischemia and infection. Current diagnostic approaches for DFU depend on comprehensive clinical assessments and imaging techniques, which are expensive and require significant resources. Traditional machine learning models often encounter challenges related to computational efficiency and precision, primarily stemming from the necessity of hand-crafted feature extraction. This study aims to evaluate the effectiveness of pretrained convolutional neural networks (CNNs), specifically GoogleNet and AlexNet, in diagnosing DFUs for three classification problems: abnormality detection (normal vs. DFU), infection, and ischemia classifications. Unlike earlier research that focused on generalizing a model for a specific task, we are working on training our models for all three datasets. A sweeping parameters technique was employed to identify the optimal training hyperparameters setting, i.e., learning rate, minibatch size, and epoch number, to enhance the model efficiency. This study further examined the model performance change with a Support Vector Machine (SVM) as the binary classifier. The results showed that the pretrained models exhibited good prediction performance with acceptable accuracy, precision, recall, and F1-score across all classification tasks ranging from 0.74-0.98. The ensemble of GoogleNet and SVM shows little to no improvement in the classification performance, while AlexNet+SVM revealed increased precision measure from 78.57 % to 82.42 % under the infection experiment. This paper concluded that further improvement of the proposed approach for DFU classification may offer a reliable solution for clinical application.