Deep learning-derived orthogonal minimum joint space width improves radiographic assessment of knee osteoarthritis severity and progression

Purpose: To evaluate the performance of a novel orthogonal minimum joint space width, defined as the narrowest distance measured orthogonally to the tibial joint line within a standardised medial axis, for detecting and monitoring knee osteoarthritis progression. This artificial intelligence-derived metric, obtained through deep learning, was compared with conventional fixed-location joint space width measures. Methods: A total of 15,313 knee radiographs from the Osteoarthritis Initiative (OAI), spanning baseline to 72-month follow-up, were analysed in this retrospective cohort study. Orthogonal minimum joint space width was automatically measured using a deep learning model. Discriminative ability for cross-sectional joint space narrowing severity was assessed using the area under the receiver operating characteristic curve. Longitudinal responsiveness was evaluated with the standardised response mean and relative standardised response mean, computed from pooled 12-month intervals and stratified by baseline narrowing grade. Variability was estimated through 1000 bootstrap resamples. Performance was compared with fixed-location measures at 22.5% and 25.0% from the medial edge of the tibial plateau. Results: Orthogonal minimum joint space width demonstrated strong discriminative performance, with area under the curve values of 0.86, 0.95 and 0.97 for joint space narrowing greater than grades 0, 1 and 2, respectively – exceeding the corresponding values for fixed-location measures by absolute margins of 0.04–0.08. Pooled relative standardised response mean analysis showed consistently higher responsiveness for orthogonal minimum joint space width, with median values of 0.97, 0.95 and 0.91 for baseline narrowing grades 0, 1 and 2, representing approximately 2%–4% greater responsiveness than fixed-location measures across all strata. Conclusions: Orthogonal minimum joint space width, automatically quantified by deep learning, enhances both sensitivity and specificity for assessing knee osteoarthritis severity and progression. This metric may serve as a robust and reproducible structural imaging endpoint for clinical trials and longitudinal research. Level of Evidence: Level II.