Applications and performance of imaging artificial intelligence in detecting and staging avascular necrosis of the femoral head: a systematic review and meta-analysis.
Authors
Affiliations (4)
Affiliations (4)
- Department of Radiology, Division of Musculoskeletal Imaging and Intervention, University of Washington, Seattle, WA, USA. [email protected].
- Department of Radiology, Division of Musculoskeletal Imaging and Intervention, University of Washington, Seattle, WA, USA.
- Department of Radiology, University Hospitals Cleveland Medical Center, Case Western Reserve University School of Medicine, Cleveland, OH, USA.
- Department of Radiology, Division of Musculoskeletal Imaging and Intervention, University of Washington, Seattle, WA, USA. [email protected].
Abstract
To evaluate the diagnostic accuracy of imaging artificial intelligence (AI) models for detecting avascular necrosis of the femoral head (AVNFH) and to examine whether diagnostic performance differs according to imaging modality, disease-stage focus, and validation strategy. In accordance with PRISMA-DTA, we conducted a systematic search in four main databases. Two independent reviewers evaluated the studies, extracted relevant data, and assessed the quality following the QUADAS-2 criteria. The primary analysis pooled overall sensitivity and specificity using a bivariate random-effects model. Prespecified subgroup analyses assessed differences by imaging modality, disease stage focus, and evaluation strategy. Nine selected estimates contributed to the primary analysis of binary AVNFH detection. Pooled sensitivity was 0.888 (95% CI, 0.789-0.944), specificity was 0.937 (95% CI, 0.906-0.958), and SROC AUC was 0.967. MRI-based estimates showed higher sensitivity than radiography-based estimates in the study-level comparison (p = 0.027). No significant subgroup differences were found by disease-stage focus and validation strategy (all p > 0.05). Only five external test-set estimates contributed to the primary analysis; subgroup comparisons had limited precision. Imaging AI models showed promising accuracy for AVNFH detection in research datasets. However, risk of bias, between-study heterogeneity, and limited external testing restrict confidence in their clinical applicability. Prospective, multicenter validation and evaluation of clinical benefit are needed before routine implementation.