Informatics and Data Science

Imaging AI for Femoral Head Osteonecrosis: PRISMA-DTA Meta-analysis of Diagnostic Accuracy and Modality Effects

Location: Concourse Foyer Kiosk

First Author(s)

PM

Peyman Mirghaderi, MD, MPH

Postdoctoral Scholar
University of Washington
Seattle, Washington, United States

Co-Author(s)

NE

Nasim Eshraghi, MD

Postdoctoral Scholar
University of Washington, United States
AA

Amir-Mohammad Asgari, MD

Student
University of Washington, United States
MP

Mohammad Taha Pahlevan Fallahy, MD

Student
University of Washington, United States
FS

Farhad Shaker, MD

Student
University of Washington, United States
MC

Majid Chalian, MD

Associate Professor
University of Washington, United States
PT

Peter C Thurlow, MD

Assisstant Professor
University of Washington, United States

Purpose: To quantify diagnostic accuracy of artificial intelligence (AI) models for detecting avascular necrosis of the femoral head (AVNFH) on imaging and assess whether performance varies by modality, disease-stage focus, and evaluation design.

Methods/Materials: A PRISMA-DTA systematic review of PubMed, Embase, Web of Science, and Scopus was performed through August 2025. Two reviewers independently screened studies, extracted 2×2 data, and assessed study quality using QUADAS-2. Sensitivity and specificity were pooled with a bivariate random-effects model (HSROC). Prespecified subgroup analyses compared modality (MRI, radiograph, CT), stage focus (early-only vs any-stage AVNFH), and evaluation strategy (train/validation, internal test, external test).

Results: Twenty-six studies met inclusion criteria.

DL-Radiomics in AVNFH Manuscri…

Fifteen detection studies contributed one best/most reliable model each (1,858 AVNFH-positive and 2,470 AVNFH-negative hips). Pooled sensitivity was 0.908 (95% CI 0.863–0.939), specificity 0.955 (0.934–0.970), and AUC 0.976 (normalized partial AUC 0.919). Leave-one-out analyses were stable and funnel-plot tests were non-significant. Modality influenced sensitivity (p=0.01): MRI-based AI achieved higher sensitivity (0.949) than radiograph-based AI (0.863), while specificity and AUC did not meaningfully differ. Disease stage focus and evaluation strategy showed no meaningful performance differences (AUC range 0.965–0.976).

Conclusions: AI models demonstrate high diagnostic performance for AVNFH detection, with higher sensitivity for MRI than radiographs. These findings support AI-enabled detection and triage workflows and define deployment priorities, including standardized reporting and broader external validation.