Artificial Intelligence Support For Digital Breast Tomosynthesis And 2DMammography: A Systematic Review Of Diagnostic Accuracy

Abstract

Background:
Artificial intelligence support for digital breast tomosynthesis and two dimensional mammography is increasingly integrated into breast cancer screening, yet the overall diagnostic accuracy and impact on workflow across clinical settings remain uncertain.
Methods:
A systematic search of PubMed from inception to May 2024 identified clinical trials and cohort studies evaluating artificial intelligence systems used as standalone readers or as decision support for radiologists in digital mammography or digital breast tomosynthesis. Studies were eligible if they reported diagnostic accuracy outcomes such as sensitivity, specificity, area under the receiver operating characteristic curve, cancer detection rate, or recall rate. Two reviewers performed screening, data extraction, and risk of bias assessment using a standardised tool, with narrative synthesis only.
Results:
Nine studies, including more than 560,000 mammographic examinations, met the inclusion criteria. Standalone artificial intelligence achieved area under the curve values around 0.84 to 0.94 and was non inferior to average radiologists in several datasets. Concurrent artificial intelligence support increased reader area under the curve by about 0.02 to 0.06, raised sensitivity by approximately 3 to 10 percentage points, and reduced reading time by 20 to 50 percent. In large screening cohorts, cancer detection rates increased by about 14 to 18 percent, while recall rates were either stable or modestly reduced and positive predictive value improved by roughly 3 to 5 percent.
Conclusions:
Artificial intelligence support for digital mammography and digital breast tomosynthesis shows consistent gains in diagnostic accuracy and efficiency, although effect sizes vary by modality, integration strategy, and case mix, and long term clinical outcomes still require confirmation.