Breast cancer leaves traces six years earlier: AI can already see them, and that forces a rethink of the entire screening process

By Zendoric · June 28, 2026.

Our thesis is straightforward: the debate over whether artificial intelligence can bring forward breast cancer detection is settled. What the study published in *Radiology*—the journal of the Radiological Society of North America (RSNA) and authored by researchers at the Karolinska University Hospital in Stockholm—demonstrates is not a laboratory promise, but a capability that is already available and wasted. The relevant question has changed in nature: it is not what AI can do, but how long health systems will take to stop ignoring what it can already do.

The data support that claim solidly. The work evaluated three commercially available AI-assisted diagnosis systems (AI-CAD), applying them retrospectively to 88,963 mammograms from 31,394 patients, drawn from the VAI-B database of four Swedish regions, with images captured between January 2008 and April 2019. Of that universe, 12,072 women were diagnosed with breast cancer during follow-up. The algorithms identified patterns consistent with the future development of the tumor in 39.3% of cases two years before the clinical diagnosis, in 25.2% four years before and—the decisive figure—in 19.7% six years before, maintaining a specificity of 90% across all time horizons. As Dr. Fredrik Strand, co-principal investigator, sums up: «Approximately 20% of breast cancer cases present mammographic signs that are already visible to AI about six years before diagnosis».

It is worth being as rigorous as the authors themselves are: we are not talking about automatic diagnoses or infallible predictions, but about probabilistic signals that would work as an early warning complementary to the radiologist's reading. AI does not replace clinical judgment; it expands it with information the human eye cannot consistently capture in routine screenings, due to structural limitations and not negligence. That distinction is no minor nuance: it is the difference between a responsible tool and an overpromise.

What distinguishes this study from others along the same lines is its time horizon: a decade of observation. Previous work had shown the usefulness of AI in estimating medium-term risk or prioritizing reviews between rounds. Extending that window to six years suggests something far more significant: mammograms contain latent information much richer than current protocols exploit. Through the lens of a trained algorithm, each image ceases to be a snapshot of the moment and becomes a point on a trajectory.

Our reading: the real change this work opens up is conceptual, not instrumental. We would move from a model of point-in-time detection—centered on the screening round—to dynamic risk monitoring over time, where each mammogram feeds an evolving individual profile and allows surveillance to be intensified in those who show rising scores, instead of applying a uniform protocol to the entire population. This connects with the underlying trend in precision oncology: personalized preventive medicine that stratifies risk continuously and adaptively, something operationally unfeasible without automation. It matters because universal screening by age, despite its proven value, carries known limitations—overdiagnosis, underdetection in dense breasts, fixed intervals—that this technology can begin to correct.

We would be naive if we presented the path as immediate, and the researchers do not. Technical questions remain to be resolved—the external validation of the algorithms in populations other than the Swedish one—as well as ethical ones—who accesses the risk scores and how they are communicated—and organizational ones—how to integrate these systems into national programs without saturating follow-up resources. The European regulatory framework for digital medical devices, which already requires robust clinical evidence for AI-CAD, adds an additional layer of rigor before any deployment at scale. None of this is trivial, but all of it is solvable.

And therein lies the argument that is hard to refute. These three systems are not prototypes: they are on the market, and their predictive capacity far exceeds what is used today in routine clinical practice. If 20% of tumors leave traces detectable by an algorithm six years before they manifest, the cost of not acting on that information has ceased to be technical or economic. It is human. The Karolinska study does not close an investigation: it opens an agenda, and the clock on that agenda is already running.