Deep learning models for osteonecrosis of the femoral head on MRI: a systematic review and meta-analysis of diagnostic performance, staging accuracy, and external generalizability.

objectivesTo evaluate the diagnostic and prognostic performance of deep learning models for osteonecrosis of the femoral head (ONFH) on MRI, including detection, staging, and collapse prediction.

methodsThis prospectively registered search was conducted in MEDLINE, Embase, Scopus, Web of Science, and IEEE Xplore. Studies applying artificial intelligence models to hip MRI for ONFH detection, staging, segmentation-assisted grading, or collapse prediction in adults were eligible. Binary staging studies and detection studies with complete 2 × 2 contingency tables were pooled using bivariate random-effects models.

resultsTwenty studies comprising 4625 patients and 6036 hips, where reported, were included. For ONFH detection, the primary bivariate meta-analysis restricted to independent observations (Scenario B; k = 4) yielded a pooled sensitivity of 0.900 (95% CI: 0.812-0.949), specificity of 0.866 (95% CI: 0.773-0.924), and HSROC AUC of 0.921; a segment-inclusive sensitivity analysis (Scenario A; k = 8 from seven unique studies) yielded 0.935 (95% CI: 0.902-0.957), 0.950 (95% CI: 0.905-0.974), and 0.971, respectively. For binary staging, 10 model-dataset pairs from six studies comprising 2593 hips were included in the diagnostic test accuracy meta-analysis; pooled sensitivity was 0.941 (95% CI: 0.901-0.965), pooled specificity was 0.877 (95% CI: 0.823-0.917), and HSROC AUC was 0.961. Leave-one-out analyses confirmed robustness, whereas external validation cohorts showed lower performance than internal or cross-validation cohorts (staging AUC, 0.895 vs 0.975; bootstrap 95% CI for the difference -0.104 to 0.009, crossing zero). For collapse prediction, evidence was limited to two studies, with external AUCs of 0.851 and 0.919 and an exploratory pooled AUC of 0.891.

conclusionDeep learning models demonstrate high diagnostic accuracy on MRI for ONFH detection and binary staging, with promising but limited evidence for collapse prediction. Because pooled staging accuracy reflects discrimination between early-stage (pre-collapse/low-grade) and late-stage (post-collapse/high-grade) disease only, and does not capture intermediate gradations within each staging system, and because performance was reduced on external validation, prospective multicenter external testing is needed before routine clinical implementation. PROSPERO REGISTRATION: CRD420261347627, Protocol version 2.0, 21 March 2026.