MRI-Based Knee Morphometry and Pediatric Anterior Cruciate Ligament Injury: A Machine Learning Case-Control Study

Background/Objectives: Anterior cruciate ligament (ACL) injury is an important cause of knee instability and long-term morbidity in children and adolescents. Although several anatomical characteristics of the knee have been associated with ACL injury, these features are commonly evaluated in isolation despite the multidimensional nature of knee morphology. This study aimed to characterize MRI-based morphometric differences associated with pediatric ACL injury and to determine whether multidimensional knee morphology could discriminate ACL-injured from control knees using machine-learning approaches. Methods: This retrospective, single-center, case-control study included 462 patients 18 years of age or younger who underwent knee magnetic resonance imaging (MRI) between January 2012 and June 2025. The cohort comprised 233 patients with ACL injury and 229 controls with an intact ACL. Eighteen MRI-derived femoral and tibial morphometric parameters were evaluated. Six feature-selection methods were combined with six classification algorithms, resulting in 36 model combinations assessed using repeated 10-fold cross-validation. Model discrimination, calibration, and feature-selection stability were evaluated. Feature selection was performed independently within each training fold to minimize information leakage. Results: In total, 13 of the 18 morphometric parameters differed significantly between groups after false discovery rate correction, including medial and lateral tibial slopes, medial tibial depth, lateral femoral condyle ratio, and intercondylar notch angles. Mean area under the receiver operating characteristic curve (AUC) values across the 36 model combinations ranged from 0.908 to 0.959. least absolute shrinkage and selection operator (LASSO) combined with naive Bayes achieved the highest mean AUC (0.959; 95% CI, 0.905-0.992), although 18 combinations showed statistically comparable discrimination. LASSO combined with logistic regression demonstrated similarly high discrimination (AUC, 0.957) with favorable calibration (Brier score, 0.080; calibration slope, 0.91; intercept, 0.01). LASSO also demonstrated the highest feature-selection stability (Nogueira index, 0.985; mean Jaccard similarity, 0.986), consistently retaining eight core features across all resampling iterations. Conclusions: Pediatric ACL injury was associated with a distinct multidimensional MRI-based morphometric profile involving both tibial and femoral anatomy. Multiple machine-learning approaches demonstrated high discrimination between ACL-injured and control knees, while LASSO-based feature selection showed particularly high reproducibility. These findings support the presence of a stable morphometric signature associated with pediatric ACL injury. The models should be interpreted as classifiers of established ACL injury status rather than prospective predictors of future injury.