Improving patellofemoral pain assessment with weight-bearing computed tomography and machine learning using three-dimensional knee joint metrics

Objectives: To evaluate whether three-dimensional (3D) knee metrics derived from weight-bearing computed tomography (WBCT) with machine learning predict patellofemoral pain severity more accurately compared with two-dimensional (2D) metrics. Methods: Diagnostic accuracy was assessed using the mean absolute error (MAE) as the primary endpoint. A five-fold cross-validation was performed for each model (random forest, gradient boosting, convolutional neural networks (CNNs), with hyperparameters tuned via grid search. The reference standard was the anterior knee pain scale (AKPS). Paired t-tests with Bonferroni correction compared with MAE differences among models. 3D knee alignment features (tilt, rotation, translations) were extracted from WBCT; 2D metrics were obtained from oblique-axial slices. Retrospective data were acquired from January to June 2022. Results: In cross-validation, random forest using 3D metrics yielded an MAE of 7.8 (95 % confidence interval (CI): 7.3–8.2), significantly lower than 8.6 (95 % CI: 8.1–9.1) in 2D-based regression (P = 0.02). CNN predictions from distal slices had an MAE of 7.5 (95 % CI: 7.0–8.0), outperforming proximal slices (8.3 (95 % CI: 7.7–8.9), P = 0.03). AKPS improved from 72 ± 10 (pretreatment) to 82 ± 6 (post-treatment) (P < 0.001). Conclusion: 3D WBCT metrics combined with machine learning significantly improved diagnostic accuracy for patellofemoral pain severity compared with conventional 2D imaging. This approach provides an objective, reproducible framework for clinical assessment and treatment planning in orthopedic practice.