Diagnostic Performance of AI-Assisted Radiologists in Breast Cancer Detection Using Digital Mammography: A Systematic...

To evaluate the diagnostic performance of AI-assisted and standalone human radiologists in breast cancer detection using digital mammography (DM). A comprehensive search was conducted in PubMed, Web of Science, Embase, Cochrane Library, and Scopus databases for studies published from January 2019 to December 2024. Study quality was assessed using the quality assessment of diagnostic accuracy studies 2 (QUADAS-2) and quality assessment of diagnostic accuracy studies-comparative (QUADAS-C). Summary receiver operating characteristic (SROC) curves and prediction regions of pooled sensitivity, specificity, and estimated area under the curves (AUCs) were used to evaluate the diagnostic performance of AI-assisted radiologists versus standalone human radiologists. Sources of heterogeneity were explored using meta-regression analysis. Overall, 30 studies were included in the qualitative synthesis. Among them, data from 20 studies were separately utilized for quantitative synthesis, categorized into three scenario groups: concurrent assistant, AI reader-replacement, and additional reader scenarios. Pooled sensitivity was significantly higher for AI-assisted radiologists compared to standalone human radiologists in the concurrent scenario (0.84 vs. 0.78, P < .001), and pooled specificity was superior in the concurrent and replacement scenarios, respectively (0.84 vs. 0.80, P < .001; 0.96 vs. 0.95, P < .001). There were no significant differences in area under the curves (AUCs) among these three scenarios. In breast cancer diagnosis, AI-assisted radiologists demonstrated superior sensitivity compared to standalone human radiologists in the concurrent scenario, and superior specificity in both the concurrent and replacement scenarios. Further research is needed to confirm these findings and explore the optimal strategies for integrating AI into breast cancer diagnostic workflows.