Artificial intelligence-based segmentation and quantitative assessment of knee cartilage in osteoarthritis: a systematic review and meta-analysis

Abstract

Purpose Quantitative MRI enables scalable assessment of knee cartilage, but the validity of artificial intelligence (AI)-derived measurements depends on more than spatial overlap. This review evaluated cartilage-segmentation accuracy and separately synthesized evidence on measurement agreement, longitudinal responsiveness, external validation, and prognostic utility.Materials and methods PubMed/MEDLINE, Embase, and Scopus were searched from inception to 1 February 2026. Dice coefficients were logit-transformed. Sampling variances were derived from reported standard deviations and test-set sizes using the delta method. The primary analysis used three-level random-effects meta-analysis with experiments nested within studies; complete cases were used when a sampling variance could be calculated. Prediction intervals and a prespecified one-estimate-per-study sensitivity analysis were reported. Analyses using imputed compartment-specific median standard deviations were secondary sensitivity analyses.Results Fifty-seven studies met the review eligibility criteria. In dependency-aware complete-case analyses of deep-learning models, pooled Dice coefficients were 0.882 (95% CI 0.854-0.905; 95% prediction interval 0.690-0.962; 41 estimates from 18 studies) for femoral cartilage, 0.869 (0.847-0.888; prediction interval 0.731-0.941; 54 estimates from 18 studies) for tibial cartilage, and 0.857 (0.837-0.875; prediction interval 0.737-0.928; 25 estimates from 15 studies) for patellar cartilage. One-estimate-per-study analyses produced similar point estimates. Only three complete-case femoral experiments from three studies were independently externally validated; their pooled estimate was imprecise (0.859, 95% CI 0.785-0.911). Reliability, longitudinal, and prognostic evidence was heterogeneous and predominantly internally validated.Conclusion AI-based knee cartilage segmentation demonstrates promising technical accuracy, but the broad prediction intervals, dependence on repeatedly used public cohorts, inconsistent reporting of dispersion, and sparse independent external validation limit generalizability. Evidence is substantially less mature for quantitative agreement, longitudinal responsiveness, and clinical prognosis than for segmentation accuracy.