In a Swedish hospital's archives sit nearly 89,000 mammograms spanning a decade—and three AI systems have just discovered something radiologists couldn't: that breast cancer often leaves detectable fingerprints years before it's officially diagnosed.
The finding, published in Radiology, matters because earlier detection has always been the holy grail of cancer care. Finding tumors when they're small and localized transforms survival rates and treatment options. But mammography has a ceiling: radiologists can only spot what's visible on the image in front of them. AI systems trained on thousands of cases can see patterns humans miss—or, more precisely, detect subtle shifts in breast tissue that later develop into cancer.
Researchers at Karolinska University Hospital in Stockholm analyzed 88,963 mammograms from 31,394 patients collected between 2008 and 2019 through Sweden's national screening program, which invites women aged 40 to 74 for examinations every two years. They ran three commercially available AI-based computer-assisted detection systems across these images, comparing cancer prediction scores over time for people who were eventually diagnosed with breast cancer against those who remained cancer-free. The results were striking: approximately 20% of breast cancer cases showed mammographic signs that were already visible to AI around six years before diagnosis.
The AI systems achieved impressive specificity—meaning they distinguished accurately between actual cancer signals and false alarms—across different time horizons. Six years before diagnosis, the AI flagged early warning signs in 19.7% of individuals who would later develop cancer while maintaining 90% specificity. That accuracy improved closer to diagnosis: four years before, it caught 25.2% of future cancers, and two years before diagnosis, it identified signs in 39.3% of cases. These aren't small numbers. Among the 31,394 patients in the study, 12,072—or 38.5%—were ultimately diagnosed with cancer by radiologists. The AI's ability to retroactively flag earlier signs in a meaningful fraction of those cases suggests a genuine opportunity to get ahead of disease progression.
Fredrik Strand, the study's senior co-author, frames the implications carefully: "Our study confirms the potential of AI to, in some cases, find signs of cancer in mammograms much earlier than when radiologists detected it." The emphasis on "in some cases" is important. This isn't about replacing human radiologists or guaranteeing earlier detection for every patient. Rather, AI could serve as an additional lens—a system that flags mammograms showing subtle changes over time, prompting radiologists to pay closer attention or recommend more frequent screening for high-risk individuals.
The Swedish approach is methodical and grounded. The national screening program's requirement that each mammogram be read by two radiologists provides rigorous ground truth. The study looked back ten years to capture cases where early signs might appear, mimicking how AI could operate in real screening settings. And the use of a large, representative database from multiple regions of Sweden strengthens the findings' applicability beyond a single hospital or demographic.
As AI tools become integrated into medical practice, studies like this one provide the evidence needed to use them wisely. Rather than replacing judgment, AI could help radiologists allocate attention where it matters most—and give more patients the chance to catch cancer when treatment options are broadest.