Automatic evaluation for knee angle measurement

Accurate measurement of the posterior tibial slope (PTS) is important for surgical planning and postoperative assessment in total knee arthroplasty and anterior cruciate ligament reconstruction. Manual methods are time-consuming and subject to variability, while automated solutions remain limited, especially for long-leg lateral radiographs. We retrospectively analyzed 816 full-length lateral radiographs from 524 patients across three centres. A deep learning framework based on a Res-UNet architecture was trained to segment anatomical landmarks and estimate the PTS. Manual measurements by experienced orthopaedic surgeons served as the reference standard. Model performance was evaluated using mean absolute error (MAE) and error distribution. Among 161 test images, the model successfully predicted PTS in 154 cases (95.7%). The overall MAE was 5.79° (range 0.01°–64.38°). Of the evaluable cases, 44.2% were within 3° and 63.0% within 5° of manual measurements. Errors greater than 10° occurred in 12.3% of cases, mainly due to contralateral limb visibility, poor image quality, or misidentification of the tibial plateau. Our findings demonstrate the feasibility of automated PTS measurement from long-leg lateral radiographs using deep learning. Although most predictions were within clinically acceptable error ranges, large outliers remain a limitation. Further refinement of landmark detection and external validation are needed. Automated methods hold promise for standardized, reproducible, and efficient PTS assessment in orthopaedic practice.