MRI-based deep learning models match or surpass physician performance for early detection and staging of avascular necrosis of the femoral head: A systematic review

Efficient and reliable magnetic resonance imaging (MRI)-based diagnosis of early avascular necrosis of the femoral head (AVNFH) is essential for guiding treatment but remains challenging due to variability in clinician experience. Deep learning (DL) models offer a promising solution. This review evaluates and summarizes the performance of DL models in the early detection and staging of AVNFH and compares their diagnostic performance with that of physicians. Three databases (PubMed, Embase, Medline [Ovid]) were searched from database inception to 25 February 2026, for articles evaluating the use of MRI-based DL models in early detection or staging of AVNFH. Studies that used only non-DL methods, did not apply DL to MRI, were not full-text, or focused on paediatric or non-human populations were excluded. Signed differences in diagnostic performance between DL models and physicians were calculated to directly compare their effectiveness. Of 1494 records, 10 studies met the inclusion criteria, comprising 1054 patients and 1958 femoral heads for early detection, 1573 patients and 2293 femoral heads for staging, and 276 patients and 533 femoral heads for both. For internal validation, average accuracy was 94.5%, sensitivity 93.2%, specificity 96.5% and AUROC 95.1%. For external validation, average accuracy was 91.4%, sensitivity 87.8%, specificity 93.3% and AUROC 91.8%. In five studies directly comparing DL models to physicians, DL models outperformed or matched physicians across most metrics, with signed differences in accuracy ranging from -10.58% to -0.43%. Among the five studies reporting statistical comparisons, the DL model significantly outperformed six of seven less experienced physicians (p&#x2009;<&#x2009;0.05), but only two of 13 experienced physicians. DL models demonstrate promising performance, performing comparably to experienced physicians while frequently outperforming less experienced physicians. Their high diagnostic accuracy and efficiency highlight their promise as valuable, objective tools to support clinical decision-making in this challenging area of musculoskeletal imaging. Level IV, systematic review.