A systematic review and meta-analysis of medical image-based artificial intelligence models for predicting metastasis...

To conduct a systematic review and meta-analysis evaluating the diagnostic performance of medical image-based artificial intelligence (AI) models for the preoperative noninvasive prediction of metastasis in ovarian cancer (OC).

We conducted a systematic literature search in PubMed, Embase, Web of Science, IEEE Xplore, Cochrane Library, CNKI, Wanfang Data, and CQVIP for studies published before February 1 ,2026. Eligibility criteria included diagnostic studies developing AI models from medical images to predict OC metastasis, with sufficient data to construct 2 × 2 contingency tables. Two reviewers independently performed data extraction and quality assessment using the QUADAS-AI tool and Radiomics Quality Score (RQS). Heterogeneity was assessed using the I² statistic. We used STATA 17.0 for meta-analysis to compute pooled sensitivity (PSen), specificity (PSpe), and area under the curve (PAUC). The protocol was registered prospectively (CRD42024619549).

9 studies involving 2343 OC patients were included. Study quality was moderate to high according to QUADAS-AI, with seven studies at low risk of bias in four or more domains. RQS ranged from 11 to 19 (median 18). The optimal model, integrating radiomics with clinical parameters, achieved a PSen of 0.81 (95%CI:0.76-0.85, I² = 5.2%), PSpe of 0.79 (95%CI:0.71-0.86, I² = 55.3%), and PAUC of 0.86 (95%CI:0.83-0.89), outperforming radiomics-only (PAUC 0.80) and clinical-only (PAUC 0.79) models. Heterogeneity for the combined model's sensitivity was low, while specificity showed moderate heterogeneity. Subgroup analyses suggested that the presence of independent validation and higher study quality (RQS ≥ 18) reduced heterogeneity. Subgroup analyses confirmed robustness across imaging modalities, algorithms, and validation methods.

Medical image-based AI models demonstrate strong performance for noninvasive prediction of OC metastasis, supporting their potential clinical utility. However, heterogeneity and variability in study quality highlight the need for more prospective, standardized, and externally validated studies.