Predicting Time-to-Total Knee Replacement: Deep Learning vs Radiomics

Motivation: Predicting time-to-total knee replacement (TKR) is essential for tailoring treatment strategies. Reliable models can identify knee osteoarthritis progression, assisting physicians in customizing care plans and potentially enhancing patient outcomes. Goal(s): Evaluate and compare the effectiveness of supervised and self-supervised deep learning (DL)-extracted features, alongside radiomics features, in predicting time-to-TKR within a 9-year timeframe. Approach: Radiomics features, and features from DL models trained in supervised and self-supervised modes were extracted from images. A Lasso feature selection was applied, and XGBoost was used for prediction. Results: Features extracted from supervised DL model achieved the highest accuracy of 71.6%, outperforming features from radiomics and self-supervised learning. Impact: This study advances predictive models for estimating time-to-TKR, aiming to support clinicians in decision-making for knee osteoarthritis management. Accurate prediction models can help patients and healthcare providers make informed treatment decisions.