JMIR Publications

ABSTRACT

Background:

Background:

Osteonecrosis of the femoral head (ONFH) is a high-incidence disable hip joint disease worldwide. Early and accurate diagnosis can significantly delay or even prevent the progression of the disease. Artificial intelligence (AI) models based on medical images have been widely studied for the diagnosis of ONFH, but there is still a lack of systematic evaluation of its diagnostic accuracy.

Objective:

Objective:

This study aims to systematically evaluate how AI models are currently used in the diagnosis of ONFH. Its diagnostic accuracy in medical imaging will also be assessed by reviewing existing research.

Methods:

Methods:

A systematic search was conducted in the PubMed, Embase, Cochrane Library, and Web of Science databases up to March 8, 2026. Studies that developed or validated image-based deep learning or machine learning models for ONFH diagnosis were included. A bivariate random-effects model was used to pool the sensitivity, specificity, positive likelihood ratio, negative likelihood ratio, and diagnostic odds ratio, and a summary receiver operating characteristic curve was plotted. Subgroup analyses were performed on the basis of imaging modality, disease stage, control group type, validation method, research center type, diagnostic criteria, and model type. Meta-regression was used to quantify the contribution of covariates to heterogeneity. Robustness and publication bias of the results were assessed using sensitivity analysis and Deek’s funnel plot asymmetry test. Fagan plots were used to assess clinical utility.

Results:

Results:

A total of 12 studies were included, covering 16,189 hip joint images. The pooled sensitivity was 0.91 (95% CI: 0.87–0.94), the pooled specificity was 0.95 (95% CI: 0.92–0.96), the pooled positive likelihood ratio was 16.7 (95% CI: 11.9–23.4), the pooled negative likelihood ratio was 0.09 (95% CI: 0.06–0.14), the pooled diagnostic odds ratio was 184 (95% CI: 101–335), and the pooled area under the receiver operating characteristic (AUC) curve was 0.97 (95% CI: 0.95–0.98). Subgroup analysis revealed that MRI-based models had significantly better diagnostic efficacy than X-ray-based models did (diagnosis odds ratio: 382 vs. 106), and the sensitivity of early ONFH models was greater than that of all ONFH models (0.93 vs. 0.89). The diagnostic odds ratio of the mixed disease control group was significantly greater than that of the healthy hip joint control group (423 vs. 111). Meta-regression revealed that imaging modality was the main source of heterogeneity, explaining 92.1% of the interstudy variance.

Conclusions:

Conclusion: AI models based on medical imaging have excellent accuracy in diagnosing ONFH, with MRI-based models performing even better and are especially suitable for early lesion detection. Future research should give priority to multi-center and forward-looking design, combined with external verification and unified diagnostic standards, explore multi-modal fusion and model interpretability, and ultimately promote the clinical transformation of AI tools. Clinical Trial: This study strictly followed the requirements of the PRISMA (Priority Reporting Items for Systematic Reviews and Meta-analyses) guidelines [24] and was registered on PROSPERO with registration number CRD420261307216.