An AI-based system for fully automated knee alignment assessment in standard AP knee radiographs

Accurate assessment of knee alignment in pre- and post-operative radiographs is crucial for knee arthroplasty planning and evaluation. Current methods rely on manual alignment assessment, which is time-consuming and error-prone. This study proposes a machine learning-based approach to fully automatically measure anatomical varus/valgus alignment in standard anteroposterior (AP) knee radiographs.

We collected a training dataset of 566 pre-operative and 457 one-year post-operative AP knee radiographs from total knee arthroplasty patients, along with a separate test set of 376 patients. The distal femur and proximal tibia/fibula were manually outlined using points to capture the knee joint. The outlines were used to develop an automatic system to locate the points. The anatomical femorotibial angle was calculated using the points, with varus/valgus defined as negative/positive deviations from zero. Fifty test images were clinically measured on two occasions by an orthopaedic surgeon. Agreement between points-based manual, automatic, and clinical measurements was assessed using intra-class correlation coefficient (ICC), mean absolute difference (MAD) and Bland-Altman analysis.

The agreement between automatic and manual measurements was excellent pre-/post-operatively with ICC 0.98/0.96 and MAD 0.8°/0.7°. The agreement between automatic and clinical measurements was excellent pre-operatively (ICC: 0.97; MAD: 1.2°) but lacked performance post-operatively (ICC: 0.78; MAD: 1.5°). The clinical intra-observer agreement was excellent pre-/post-operatively with ICC 0.99/0.95 and MAD 0.9°/0.8°.

The developed system demonstrates high reliability in automatically measuring varus/valgus alignment pre- and post-operatively, and shows excellent agreement with clinical measurements pre-operatively. It provides a promising approach for automating the measurement of anatomical alignment.