Artificial intelligence for ovarian cancer diagnosis via ultrasound: a systematic review and quantitative assessment of model performance

TL;DR

This paper reviews AI models for diagnosing ovarian cancer via ultrasound, finding high accuracy but highlighting the need for better validation and standardization.

Contribution

The study provides a systematic review and meta-analysis of AI-based ultrasound diagnostics for ovarian cancer, identifying key performance metrics and methodological challenges.

Findings

AI models achieved high pooled accuracy (92.3%), sensitivity (91.6%), and AUC (0.93) for ovarian cancer diagnosis.

Automated segmentation outperformed manual segmentation in diagnostic accuracy and sensitivity.

Methodological rigor, not dataset size, was the primary determinant of model performance.

Abstract

Early and accurate detection of ovarian cancer (OC) remains clinically challenging, prompting exploration of artificial intelligence (AI)-based ultrasound diagnostics. This systematic review and meta-analysis critically evaluate diagnostic accuracy, methodological rigor, and clinical applicability of AI models for ovarian mass classification using B-mode ultrasound.

A systematic literature search following PRISMA guidelines was conducted in PubMed, IEEE Xplore, and Scopus up to December 2024. Eligible studies included AI-based ovarian mass classification using B-mode ultrasound, reporting accuracy, sensitivity, specificity, and/or area under the ROC curve (AUC). Data extraction, quality assessment (PROBAST), and meta-analysis (random effects) were independently performed by two reviewers. Heterogeneity sources were explored.

From 823 identified records, 44 studies met inclusion criteria, covering over 650,000 images. Pooled performance metrics indicated high accuracy (92.3%), sensitivity (91.6%), specificity (90.1%), and AUC (0.93). Automated segmentation significantly outperformed manual segmentation in accuracy and sensitivity, demonstrating standardization benefits and reduced observer variability. Dataset size minimally correlated with performance, highlighting methodological rigor as a primary determinant. No specific AI architecture consistently outperformed others. Substantial methodological heterogeneity and frequent risk-of-bias issues (limited validation, small datasets) currently limit clinical translation.

AI models show promising diagnostic performance for OC ultrasound imaging. However, addressing methodological challenges, including rigorous validation, standardized reporting (TRIPOD-AI, STARD-AI), and prospective multicenter studies, is essential for clinical integration. This review provides clear recommendations to enhance clinical translation of AI-based ultrasound diagnostics.

B-mode ultrasound images show ovarian tumors in original, manual, and automatic segmentation. A comparison between expert diagnosis and artificial intelligence highlights AI's accuracy, sensitivity, specificity, and AUC. Illustrations depict healthy, early, and advanced tumor stages, with AI symbols like a magnifying glass, network, and brain.

Figures