One of the largest real-world tests of artificial intelligence in cancer screening to date, a nationwide German study known as PRAIM, found that radiologists using an AI-assisted mammography viewer detected breast cancer at a rate roughly 17.6% higher than those reading scans without it — without an increase in false positives. The study, led by cancer epidemiologist Alexander Katalinic of the University of Lübeck and published in Nature Medicine, followed close to 500,000 women across Germany’s national screening program, with more than 460,000 women aged 50 to 69 contributing data between July 2021 and February 2023 across 12 screening sites and 119 participating radiologists.
Why Mammography Needed Help in the First Place
Reading mammograms is widely considered one of radiology’s most demanding visual tasks: tumors can be small, subtle, and easily obscured by normal variations in breast tissue density, and studies going back decades have documented meaningful rates of both missed cancers and unnecessary callbacks even among experienced double-reading radiologists, the standard practice in many European screening programs. That persistent error rate — in both directions — is precisely what made breast screening one of the earliest and most heavily researched applications for AI in medical imaging, starting with retrospective studies years before PRAIM attempted to test the technology at genuine national scale rather than in a curated research dataset.
What Made PRAIM Different From Earlier AI Mammography Studies
Unlike many AI-in-radiology studies that compare algorithm performance against a fixed, retrospective set of images under laboratory conditions, PRAIM was a prospective cohort study embedded directly in Germany’s ongoing, nationwide mammography screening infrastructure, with the AI-supported viewer used for roughly half of all screenings conducted during the study period. That design meant the AI tool had to perform under the same conditions — varying equipment, varying radiologist experience, varying patient populations — that any national rollout would actually face, rather than the more favorable conditions of a controlled research trial.
The Case That This Settles the Debate
Katalinic has argued the results should move the conversation past whether AI belongs in screening mammography at all: “AI in mammography screening is at least as good as a human reader, and our study shows it’s even better,” he said, pointing to the roughly one additional cancer detected per 1,000 women screened without any corresponding rise in false positives — a combination that, if it holds up in other health systems, would represent a rare case of a screening improvement with no apparent tradeoff. Stefan Bunk, chief technology officer and cofounder of Vara, the Berlin-based company that developed the AI system used in the study, said the results should shift the conversation toward how AI fits into radiologists’ daily workflow rather than whether it belongs there.
What the Study Doesn’t Settle
A roughly 17.6% relative increase in detection, while statistically significant across a half-million-patient dataset, still represents a modest absolute difference per individual screening, and the study’s design — AI used as an assistive second reader embedded in an existing double-reading system — does not necessarily generalize to countries like the United States, where single-reader mammography is the norm and the AI tool would be replacing a second human reader entirely rather than supplementing one. The study also did not report on longer-term outcomes such as interval cancer rates or mortality differences, the gold-standard measures that determine whether earlier detection actually translates into saved lives rather than simply finding cancers that would have been caught at the next scheduled screening regardless.
What’s Next
PRAIM’s scale makes it one of the most-cited pieces of evidence as health systems beyond Germany debate how, and how fast, to fold AI into national screening programs — a conversation already underway in the UK’s own AI Diagnostic Fund and in ongoing U.S. FDA clearances for mammography AI tools. The next wave of research, including longer-term outcome tracking and studies in single-reader screening environments, will determine whether the detection gains documented in Germany’s double-reading system hold up as other countries adapt the technology to their own, often very different, screening infrastructure.
Photo: Mitrey / PIXABAY via Pixabay
