Diagnostic performance of machine learning algorithms in breast ultrasound: A systematic review and meta-analysis

Abstract

The diagnosis of breast neoplasms by ultrasound examination depends on the subjective interpretation of the radiologist, which determines the need to develop objective methods of computer analysis to standardise the diagnostic process. The aim of this study was to determine the diagnostic performance of machine learning algorithms for classifying benign and malignant lesions in breast ultrasound through a systematic literature review with quantitative meta-analysis. A systematic search was conducted in PubMed, Scopus and Embase in accordance with the Preferred Reporting Items for Systematic Reviews and Meta-Analyses guidelines, identifying 57 studies published between 2019 and 2025, of which 49 were included in the meta-analysis using a bivariate model and quality assessment with the Quality Assessment of Diagnostic Accuracy Studies tool. The pooled sensitivity was 91.1% (95% confidence interval 88.9-93.0%), specificity 85.3% (95% confidence interval 82.9-87.5%), area under the receiver operating characteristic curve 0.957 (95% confidence interval 0.941-0.973), and diagnostic odds ratio 64.8 (95% confidence interval 46.1-91.2). Subgroup analysis showed that deep learning demonstrated higher specificity compared with commercial systems (87.2% versus 81.5%, p=0.041), a multimodal approach provided a 3.9% gain (p<0.001), artificial intelligence assistance was most effective for junior radiologists with an improvement of 0.164 (p<0.001), and use for reclassification of BI-RADS 4A lesions allowed biopsy reduction of 67.8-79.1% with a negative predictive value of 96.7%. The findings support clinical implementation of artificial intelligence systems for standardising diagnosis and optimising biopsy indications. The practical significance lies in reducing unnecessary invasive procedures and improving the diagnostic performance of less experienced radiologists.