Ultrasound-Based Deep Learning Models Performance versus Expert Subjective Assessment for Discriminating Adnexal Masses: A Head-to-Head Systematic Review and Meta-Analysis

Abstract

(1) Background: Accurate preoperative diagnosis of ovarian masses is crucial for optimal treatment and postoperative outcomes. Transvaginal ultrasound is the gold standard, but its accuracy depends on operator skill and technology. In the absence of expert imaging, pattern-based approaches have been proposed. The integration of artificial intelligence, specifically deep learning (DL), shows promise in improving diagnostic precision for adnexal masses. Our meta-analysis aims to evaluate DL’s performance compared to expert evaluation in diagnosing adnexal masses using ultrasound images. (2) Methods: Studies published between 2000 and 2023 were searched in PubMed, Scopus, Cochrane and Web of Science. The study quality was assessed using QUADAS-2 (Quality Assessment of Diagnostic Accuracy Studies-2). Pooled sensitivity and specificity for both methods were estimated and compared. (3) Results: From 1659 citations, we selected four studies to include in this meta-analysis. The mean prevalence of ovarian cancer was 30.6%. The quality of the studies was good with low risk of bias for index and reference tests, but with high risk of bias for patient selection domain. Pooled sensitivity and specificity were 86.0% and 90.0% for DL and 86.0% and 89.0% for expert accuracy (p = 0.9883). (4) Conclusion: We found no significant differences between DL systems and expert evaluations in detecting and differentially diagnosing adnexal masses using ultrasound images.

Keywords

• adnexal mass
• deep learning
• transvaginal ultrasound
• ovarian cancer
• artificial intelligence