Predicting Rapid Progression in Knee Osteoarthritis: A Novel and Interpretable Automated Deep Learning Approach Using DenseNet, with Specific Focus on Young Patients and Early Disease

Abstract

Knee osteoarthritis (KOA) is a degenerative condition that leads to the destruction of cartilage and joint tissues. Identifying patients who show rapid progression is really important for starting treatment early and preventing early joint damage. Here, we propose a deep learning approach that is interpretable and based on the DenseNet 121 architecture, which can predict KOA progression after obtaining radiographic knee images at baseline. The design of the model is such that concentrating on the disease at an early stage and in younger individuals is possible, so then, a step ahead can be taken in effective management of the patients and also in the provision of care that is personalized care. Through DenseNet’s densely connected layers, efficient reuse of heavy features and comprehensive depiction of very fine radiographic degenerative patterns are achievable. The system built was given the OAI data as input for training and was externally evaluated using the MOST data. Rapid progression is considered an increase of two or more KL grades within a 24 48 month timescale. The model proposed displayed a top level of prediction performance (AUC = 0.86, C index = 0.83), and it was also able to properly generalize external data (AUC = 0.84). Visual interpretability methods, such as Grad CAM and Integrated Gradients, in our study also identified the anatomical features, i.e., joint space narrowing and osteophyte formation associated with the disease, which are the key evidence of the disease, thereby strengthening the model's transparency. Results indicate that the proposed method based on a DenseNet should serve as an accurate and explainable model to predict KOA progression at an early stage, leading to more effective and personalized disease management.