Detection of diabetic foot ulcers using artificial intelligence

Abstract

Diabetic foot ulcers (DFUs) are one of the most common and devastating complications of diabetes, representing a considerable challenge for healthcare systems and significantly affecting patients' quality of life. This paper is aimed to address the application of artificial intelligence techniques for DFU detection. Deep learning (DL) models were implemented, specifically YOLOv8, in its m and l variants, and standard and deformable convolutions Faster R-CNN. These models were optimized using data augmentation and hyperparameter tuning techniques. The dataset provided by the Diabetic Foot Ulcers Grand Challenge 2020 (DFUC 2020) was used for training. Additionally, a national image database was used to evaluate the model's effectiveness in the local context. In the training and validation stages, mAP greater than 0.70 was obtained for all trained models, being the Yolov8l model the best performed (mAP=0.799). The external test with national data demonstrated generalization power, granting accuracy values ​​of 0.856, sensitivity of 0.789, F1 score of 0.821, and an mAP50 of 0.839. A web application was also developed using Streamlit, providing an interactive interface for analyzing DFU images. The object detection DL models implemented proved to be effective tools for accurately identifying DFUs. The developed application effectively integrates the best performed model, offering an intuitive interface for healthcare professionals.