Comprehensive Summary
This study presents an AI system used to automatically assess knee alignment from anteroposterior (AP) knee radiographs in patients undergoing total knee arthroplasty (TKA). Researchers trained machine learning models (MLMs) on 1,023 manually outlined pre- and post-operative AP knee radiographs. The system calculated the anatomical femorotibial angle (aFTA) and defined varus and valgus deformities as negative and positive deviations from 0, respectively. Its performance was validated on a separate test set of 376 patients by comparing AI results to manual and clinical measurements, using intra-class correlation coefficient (ICC), mean absolute difference (MAD), and Bland-Altman analysis. The AI demonstrated excellent agreement with manual measurements both pre- and post-operatively, and strong agreement with clinical measurements pre-operatively. However, agreement with post-operative clinical measurements was lower. Increased performance was achieved when specific anatomical features such as the femoral notch and tibial spines were used to define aFTA. Errors were associated with previous fractures and rotated or shortened femoral and tibial shafts. While the AI model can reliably perform manual and clinical pre-operative assessments, its post-operative performance needs further improvement.
Outcomes and Implications
Accurate assessment of knee alignment is needed for planning and evaluating total knee arthroplasty (TKA), as misalignment can lead to adverse patient outcomes. However, current manual methods are time-consuming and can be potentially variable. An AI-based system offers a reliable and fully automated solution for measuring knee alignment from standard AP knee radiographs. The system has a strong potential for integration into pre- and post-operative assessments. Although not yet in clinical use, the authors highlight that the software can run on standard hardware and is publicly available for further validation, helping accelerate its use in clinical practice.