What question did this study set out to answer?

This work aims to assess the diagnostic accuracy and clinical utility of artificial intelligence in breast imaging modalities.

May 24, 2026Open Access

404eP Clinicopathological profile and outcomes of premenopausal women with breast cancer carrying germline pathogenic variants: a real-world cohort from a Northern Portugal cancer center

Key Points

This work aims to assess the diagnostic accuracy and clinical utility of artificial intelligence in breast imaging modalities.
Systematic search of PubMed, Scopus, and CENTRAL for meta-analyses from 2015-2025 on AI applications in breast imaging.
Inclusion of 18 meta-analyses evaluating mammography, ultrasound, and MRI for diagnostic assessment.
Quality assessment of reviews using AMSTAR 2 framework.
Evidence supports AI as an adjunct to mammography, with performance within radiologists' range but varying specificity.
About 2/3 of examinations can be deprioritized while maintaining high sensitivity; results depend on specific algorithms and thresholds.
Major limitations were noted in methodological quality with high confidence in 1 review, low in 4, and critically low in 13.

Abstract

Background: Breast imaging services face increasing examination volumes and variability in reader experience, creating pressure to maintain cancer detection while limiting unnecessary recalls.Artificial intelligence (AI) has been proposed to support screening and diagnostic workflows across breast imaging modalities.Methods: PubMed, Scopus, CENTRAL, and major radiology and oncology congress proceedings were searched (2015-2025) for meta-analyses (MAs) evaluating AI applied to mammography/digital breast tomosynthesis (DBT), ultrasound (US), or magnetic resonance imaging (MRI) for screening and/or diagnostic assessment.Eligible MAs synthesised diagnostic accuracy, reader comparison, AI assistance, or triage strategies.Two reviewers independently screened records, extracted data, and assessed review quality using AMSTAR 2. Evidence was summarised by modality and clinical use-case.Results: From 586 studies, 18 MAs were included.Evidence was most consistent for screening mammography/DBT.Head-to-head syntheses generally placed standalone AI within the performance range of radiologists, although specificity varied across operating thresholds and study designs.MAs of AI assistance or second reading suggested improved decision support and consistency, but comparators and access to prior examinations were often incompletely reported.Triage-focused syntheses indicated that approximately 2/3 of examinations might be safely deprioritised while maintaining high sensitivity, although findings were algorithm-and threshold-specific and programme-level outcomes such as interval cancers were inconsistently assessed.US and MRI meta-analyses demonstrated promising discrimination but were dominated by retrospective, single-centre cohorts with heterogeneous protocols and limited external validation.Overall AMSTAR 2 confidence was high in 1 review, low in 4, and critically low in 13.Conclusions: Current meta-analytic evidence supports AI as an adjunct to screening mammography/DBT workflows, but evidence for breast US or MRI is still advancing.Most MAs have methodological limitations, so headline pooled estimates should be interpreted cautiously.

Mark Helpful

Bookmark

Relay

View Full Paper