INTRODUCTION: Delays of nearly a decade are common in endometriosis diagnosis, leaving women with unmanaged pain, impaired fertility, and repeated healthcare visits. Although diagnostic imaging has advanced, confirmation still requires invasive laparoscopy, limiting timely diagnosis. Ultrasound is widely available and cost-effective, but its accuracy varies because it depends heavily on operator expertise, leading to inconsistent performance. Artificial intelligence (AI) offers the potential to transform ultrasound into a reproducible, scalable, and noninvasive diagnostic tool. By using deep learning to analyze sub-visual radiomic features from standard ultrasound images, AI models could improve diagnostic accuracy. Several AI models have been proposed, but their performance has not been systematically evaluated across different disease subtypes, input features, or algorithm types. This meta-analysis synthesizes existing evidence to assess the diagnostic accuracy of AI-driven ultrasound radiomics, exploring its potential as a non-invasive, reliable tool for patient triage. Ultimately, AI could reduce reliance on surgery and bridge the diagnostic gap in endometriosis care. OBJECTIVE: To conduct the first comprehensive systematic review and meta-analysis of AI-based ultrasound radiomics models for endometriosis detection and to identify which AI models' characteristics yield optimal performance. METHODS: A PRISMA-guided review of PubMed, Web of Science, Scopus, EBSCO, and IEEE Xplore was conducted for studies from inception to September 2025. Included studies evaluated AI-based ultrasound diagnosis against surgical reference standards, with histologic confirmation when available. Data on study design, patient populations, model types, input modalities, and outcomes were extracted. Pooled sensitivity, specificity, and AUC were calculated using univariate and bivariate models. Subgroup analyses examined performance across endometriosis subtypes, input features, algorithm families, and ultrasound modalities (PROSPERO: CRD420251156983). RESULTS: The analysis included 13 datasets from 11 studies with 2,727 patients (Table 1). Overall, AI models showed high diagnostic accuracy: sensitivity of 87.2% (95% CI: 80.8–91.7), specificity of 89.9% (95% CI: 82.8–94.3), and an AUC of 0.93. Subgroup analysis revealed higher sensitivity for endometrioma (93.5%) versus deep infiltrating endometriosis (82.3%) (p=0.02) (Figure 1). Radiomics-only models outperformed combined radiomics and clinical data models (90.1%/92.6% vs 75.4%/77.5%) (p<0.01) (Figure 2). The QUADAS-2 tool indicated low risk of bias in most studies. No significant publication bias was found (p=0.15). Heterogeneity across studies was notable, reflecting variability in methodologies and technical factors. CONCLUSIONS: AI-based ultrasound demonstrates high diagnostic accuracy for endometriosis, with better performance for endometrioma than deep infiltrating disease. Radiomics-only models outperform combined approaches, supporting AI as a noninvasive adjunct that can standardize ultrasound interpretation, reduce operator dependency, and help close the diagnostic gap. This could optimize resource use, improve health equity, and reduce delayed diagnoses. However, significant heterogeneity from differences in study design, patient selection, imaging protocols, and equipment highlights the need for rigorous, multicenter prospective studies with standardized methodologies, device reporting, and external validation to ensure reproducibility and clinical adoption.Figure 1Figure 2Table 1
Ghanbari et al. (Fri,) studied this question.