Text

Objectives and questions: Axial deformities of the lower limbs in adolescents and young adults are clinically relevant due to their association with progressive joint degeneration, patellar instability and functional impairment. Precise assessment of lower limb alignment is essential for diagnosis, preoperative planning and postoperative monitoring. Standing long-leg radiographs are the clinical standard for evaluating limb alignment, requiring precise angle and length measurements. Recent advances in artificial intelligence (AI) offer the potential for automated, rapid and reproducible measurement, which could reduce the clinical workload. This study aimed to compare software-assisted manual measurements with fully automated, AI-based analysis.

Material and methods: This retrospective study included adolescents with closed growth plates and adults up to 40 years of age who were referred between 2013 and 2023. Of the 4,770 legs screened, strict radiographic quality criteria resulted in a final cohort of 304 legs. Most cases were non-operative and pathology-free. The most common pathological findings were standing blocks, metal implants, followed by tumors and fractures. Radiographs were analysed manually using TraumaCAD (Brainlab, Munich, Germany) and by AI software BoneMetrics (Gleamer, Paris, France). The comparison of both measurements included the following parameters: mechanical Lateral Proximal Femoral Angle (mLPFA), mechanical Lateral Distal Femoral Angle (mLDFA), mechanical Medial Proximal Tibial Angle (mMPTA), mechanical Lateral Distal Tibial Angle (mLDTA), joint line convergence angle (JLCA), Mechanical Axis Deviation (MAD), femur length, tibial length and total leg length. Data analysis included Bland-Altman plots, intraclass correlation coefficient (ICC) regression analyses.

Results: The mean bias for angular measurements ranged from −0.48° to 0.57°, indicating no systematic over- or underestimation. Length measurements showed a small negative bias, ranging from −47.38 mm to −0.05 mm, suggesting slight underestimation. Overall, AI-based measurements demonstrated clinically acceptable agreement with manual analysis. Angular measurements were more reliable than length measurements, likely due to dependency on calibration. ICC analysis revealed moderate to excellent agreement (ICC>0.72) for all parameters, except for the JLCA, which demonstrated poor reliability (ICC=0.15).

Discussion and conclusions: Overall, AI-based measurements showed clinically acceptable agreement with manual measurements, with angular measurements demonstrating higher reliability than bone length measurements. This difference may be attributable to calibration dependency affecting length assessment. The comparatively poor reliability of AI-based JLCA measurements may reflect inherent uncertainty or systematic errors. This finding is of particular clinical relevance, as JLCA has prognostic significance in corrective osteotomies. While AI-based analysis may serve as a support tool in clinical practice, caution is required when interpreting JLCA values.