Artificial intelligence for breast cancer screening in mammography (AI-STREAM): preliminary analysis of a prospective multicenter cohort study
Comprehensive Overview
Published in Nature Communications in 2025, this study asked a practical question facing breast cancer screening programs worldwide: does adding an artificial intelligence (AI) assistant to a radiologist's routine mammogram reading actually catch more real cancers, and if so, does it come at the cost of more unnecessary follow-up visits for healthy women? Researchers followed 24,543 women aged 40 and older who underwent mammograms at six hospitals as part of South Korea's national breast cancer screening program between 2021 and 2022. Each mammogram was first read by an experienced breast radiologist working alone, then reviewed a second time with the help of a commercial AI program that highlights suspicious areas and scores each breast for cancer likelihood. Radiologists made their final decision about whether to call a woman back for further testing only after seeing both reads. The researchers then tracked which women were later confirmed, through biopsy and pathology, to actually have cancer, comparing how many cancers were caught, and how many women were recalled unnecessarily, with versus without the AI's input. As expected from smaller prior studies, adding AI meaningfully improved cancer detection without significantly increasing how often healthy women were called back for extra testing.
Outcomes and Implications
The headline result is that AI assistance let breast radiologists find nearly 14% more cancers than they caught alone, 140 cancers with AI-assisted reading versus 123 without it, out of roughly 24,500 women screened, while the share of women called back for further testing barely budged. The benefit was even larger for general radiologists without specialized breast-imaging training, whose cancer detection jumped by about 26% with AI help, though unlike the breast specialists, these general radiologists also recalled meaningfully more women unnecessarily once AI was added, suggesting the technology helps less experienced readers most but also makes them more trigger-happy. Notably, the AI system used alone, with no radiologist involved, performed no better than the specialist radiologists but generated significantly more unnecessary recalls, underscoring that AI currently works best as a second opinion rather than a replacement. Cancers caught with AI's help tended to be smaller, less likely to have spread to lymph nodes, and of more treatable molecular subtypes, hinting that AI may catch cancers earlier in their course. Current mammography screening already reduces breast cancer deaths substantially but relies heavily on radiologists' judgment, and busy screening programs worldwide face persistent radiologist shortages. These findings suggest AI support could meaningfully strengthen real-world screening accuracy, particularly for less experienced readers, without the tradeoff of excess false alarms. The authors caution these are preliminary, observational results, and conclude longer follow-up is needed to confirm the gains translate into fewer missed cancers.