Privacy-preserving knee osteoarthritis classification: A federated learning approach with GradCAM visualization

Abstract

Knee Osteoarthritis (KOA) poses a significant global health challenge, impacting a substantial population. The conventional detection process involves multiple tests and meticulous examination by experienced physicians, which is time-consuming and susceptible to misclassification due to subtle variations in X-ray images. Additionally, privacy concerns hinder the sharing of sensitive data like X-rays. This study employs Federated Learning with pre-trained architectures (DenseNet-169, Inception-v2, and MobileNet-v2) to classify three KOA severity grades, utilizing two clients to ensure data privacy. The aim is to develop a generalized model for disease classification, improving efficiency while ensuring the confidentiality of patient information. DenseNet-169 excelled with an F1 score of 81% and an accuracy of 82%, while Inception-v2 and MobileNet-v2 performed well with slight F1 score variations. Moreover, the exploration of GradCAM visualization techniques is conducted to improve interpretability, highlighting the capability of this approach to effectively tackle the intricate challenges linked to Knee Osteoarthritis detection.