Machine and deep learning models for ligament injury recognition: a systematic review and meta-analysis of imaging and novel diagnostic techniques

Purpose: Diagnosing ligament injuries remains a challenge for musculoskeletal clinicians due to the lack of standardized classification, evaluation, and management protocols. Machine learning (ML) and deep learning (DL) models offer potential to improve diagnostic accuracy. This study aimed to evaluate the diagnostic performance of various ML and DL models in identifying ligament injuries across different medical imaging modalities. Methods: A meta-analysis was conducted following the PRISMA 2020 checklist. Searches were performed in PubMed, SCOPUS, Web of Science, and the Cochrane Library. Study quality was assessed using the QUADAS-2 tool and Robvis software. Diagnostic performance measures - true positive, true negative, false positive, and false negative - were analyzed. A random-effects model was applied, and heterogeneity and subgroup analyses were conducted. Statistical and graphical analyses were performed using R. The study was registered in PROSPERO (CRD42025646317). Results: Fifty-nine ML and DL algorithms from 23 studies were analyzed. Pooled sensitivity and specificity were 0.890 (95% CI: 0.829-0.938) and 0.926 (95% CI: 0.820-0.959), respectively. Pooled estimates for PLR, NLR, lnDOR, and AUC were 1,644.37 (95% CI: 73.56-3,215.18), 0.179 (95% CI: 0.095-0.263), 4.130 (95% CI: 3.570-4.700), and 95%, respectively, with P < 0.001. Conclusion: ML and DL models demonstrate high diagnostic accuracy in detecting ligament injuries. Their strong performance supports ongoing integration into clinical practice, offering valuable support for musculoskeletal specialists in image interpretation and diagnosis.