Radiomics-Based Machine Learning for the Detection of Myometrial Invasion in Endometrial Cancer: Systematic Review and Meta-Analysis.

BACKGROUND

Preoperative endometrial cancer (EC) diagnosis often depends on radiologists' expertise, which introduces subjectivity. Recent studies have explored radiomics-based machine learning (ML) models for detecting myometrial invasion (MI), but a comprehensive evaluation of their diagnostic performance is lacking. Therefore, our study systematically assessed the diagnostic performance of radiomics-based ML approaches for identifying MI in EC, thereby providing evidence to guide the development or improvement of noninvasive diagnostic tools.

OBJECTIVE

This study aims to systematically assess the diagnostic performance of radiomics-based ML approaches for identifying MI in EC and compare the diagnostic efficacy of conventional ML (CML) and deep learning (DL) models based on differences in data processing methods via subgroup analyses, thereby providing evidence to guide the development or improvement of noninvasive diagnostic tools.

METHODS

PubMed, Cochrane Library, Embase, and Web of Science were searched through November 26, 2024, for studies evaluating radiomics-based ML for detecting MI in patients with EC. Study quality was appraised using the radiomics quality score. Pooled diagnostic metrics were estimated using a bivariate random-effects model. Subgroup analyses compared CML and DL models.

RESULTS

We included 19 studies comprising 4373 patients with EC. Of these 19 studies, 18 (95%) used magnetic resonance imaging-based radiomics, and 1 (5%) used ultrasound imaging. The pooled estimates from the meta-analysis demonstrated a sensitivity of 0.79 (95% CI 0.73-0.83), a specificity of 0.83 (95% CI 0.79-0.86), a positive likelihood ratio (PLR) of 4.5 (95% CI 3.5-5.8), a negative likelihood ratio (NLR) of 0.26 (95% CI 0.20-0.34), a diagnostic odds ratio (DOR) of 17 (95% CI 11-28), and an area under the summary receiver operating characteristic curve (AUSROC) of 0.89 (95% CI 0.00-1.00). Subgroup analyses revealed that the DL models achieved a sensitivity of 0.81 (95% CI 0.71-0.88) and a specificity of 0.86 (95% CI 0.76-0.92). The PLR, NLR, DOR, and AUSROC were 5.6 (95% CI 3.2-9.8), 0.22 (95% CI 0.14-0.36), 25 (95% CI 10-64), and 0.89 (95% CI 0.00-1.00), respectively. By contrast, the CML models exhibited a sensitivity of 0.77 (95% CI 0.69-0.83) and a specificity of 0.81 (95% CI 0.77-0.85). The PLR, NLR, DOR, and AUSROC were 4.1 (95% CI 3.2-5.4), 0.28 (95% CI 0.20-0.39), 15 (95% CI 9-25), and 0.86 (95% CI 0.00-1.00), respectively.

CONCLUSIONS

Radiomics-based ML shows strong potential for noninvasive prediction of MI in EC, with DL outperforming CML. However, current evidence is limited and relies mainly on internal validation. Larger-scale, multicenter studies are needed to establish robust artificial intelligence-based diagnostic tools.

TRIAL REGISTRATION: PROSPERO CRD420250625797; https://www.crd.york.ac.uk/PROSPERO/view/CRD420250625797.