Identification and Classification of Diabetic Foot Ulcers by Deep Neural Networks: A Comparative Study

Automatic analysis of the Diabetic Foot Ulcer (DFU) wound using computer based methods is becoming important with the rapid development of image-based machine learning and deep learning algorithms. In this work, identification and classification of diabetic wounds have been studied utilizing several deep neural network models, and their performances have been compared. Simulation experiments have been done utilizing the DFUC2021 data set, containing labeled images of four classes: normal class, infection only class, ischemia only class, both (infection and ischemia) class. The final dataset of DFUC2021 comprises 15,683 DFU images in total, with 5955 training images, 5734 testing images, and 3994 unlabeled DFU patches. A few deep network model architectures, such as VGG16, VGG19, ResNet50, ResNet101, and EfficientNetB0 which were pretrained have been utilized for the study. Initially, the original data set was used for training and classification in which the classes are not balanced. Data augmentation has been utilized as a means of oversampling to equalize all the samples in all the classes. The performance study has been done by comparing the values of precision, accuracy, recall, F1 score values, and computational time for the networks utilizing original and augmented datasets and a comparative analysis is reported.