Can AI Spot Cancer Earlier? Where Imaging Helps and Where False Alarms Matter

In a breast-screening clinic, the images arrive in a steady stream: four grey views per woman, thousands of women per week. For decades, many European programs have had each mammogram read by two radiologists, one after the other, to catch what a single reader might miss. Now, in some of those programs, one of the first “readers” to see the image is software.

That change raises a question with two halves. Can artificial intelligence help find cancers earlier? And what happens to everyone else — the far larger number of people who don’t have cancer but might be flagged anyway? The best evidence so far comes from breast screening, and it is more encouraging than most AI health claims. It also shows why “finds more cancer” is only the start of the story.

Why “earlier” is not automatically “better”

Screening means testing people who have no symptoms. Because almost everyone screened is healthy, even a small error rate touches many lives. NCI’s expert summary on breast cancer screening sets out the trade-offs that any AI tool inherits:

• False positives — a result that looks suspicious but turns out not to be cancer. NCI estimates that approximately half of U.S. women screened every year for ten years will experience at least one, and some of those will undergo a biopsy.
• False negatives — a cancer that is present but missed. Mammography misses more cancers in dense breast tissue and certain tumor types.
• Overdiagnosis — finding a real cancer that would never have caused symptoms or harm in the person’s lifetime. Overdiagnosed cancers still get treated, with all the side effects that brings, and can’t be told apart from dangerous ones at the time of diagnosis.

False alarms are not only a moment of fear. A large NCI-funded study of 3.5 million mammograms, described on NCI’s Cancer Currents blog, found that women who received a false-positive result were less likely to come back for their next routine screen: 77% returned after a clean result, compared with 61% of those called back for extra imaging and 67% of those who had a biopsy. A tool that raises alarms carelessly could, in principle, make screening less effective over time.

Plain English: Sensitivity is how many real cancers a test catches. Specificity is how many healthy people it correctly clears. Good screening needs both — improving one at the expense of the other just moves the harm around.

The strongest evidence: a randomized trial in Sweden

The Mammography Screening with Artificial Intelligence trial, known as MASAI, is the kind of study the field has long needed: a randomized controlled trial, in which women were assigned by chance to either AI-supported screening or standard double reading by two radiologists. More than 100,000 women in Sweden took part between April 2021 and December 2022. In the AI arm, software helped sort exams by estimated risk and highlight suspicious areas, but at least one human radiologist still read every exam.

An earlier analysis reported a 44% reduction in radiologists’ screen-reading workload. The full results, published in The Lancet in early 2026, looked at what matters more: cancers that surface between screening rounds, known as interval cancers, which are often the ones screening missed.

• Interval cancers occurred at a rate of 1.55 per 1,000 women in the AI group and 1.76 per 1,000 in the control group. The trial was designed to show AI-supported screening was not worse, and it met that goal; the difference itself was not statistically significant.
• Sensitivity was higher with AI support: 80.5% versus 73.8%.
• Specificity was the same in both groups, at 98.5% — meaning more cancers were caught without more false alarms.

The trial also reported fewer interval cancers with aggressive or advanced features in the AI arm. But the investigators, in the Lancet’s announcement, spelled out the limits: one country, one type of mammography machine, one AI system, moderately to highly experienced radiologists, and no data collected on race or ethnicity. Results may not transfer to settings that differ on any of those points.

An AI that finds more cancers is only half an answer. The other half is how many healthy people it worries along the way — and whether the extra cancers are ones that would have caused harm.

Real-world evidence: Germany’s PRAIM study

Trials show what can happen under controlled conditions; real-world studies show what does happen in routine care. In Germany’s national screening program, the PRAIM study followed more than 460,000 women screened at 12 sites between 2021 and 2023. The study compared exams read with AI support against those read by standard double reading alone. According to the University of Lübeck’s announcement of the results, published in Nature Medicine in January 2025, the AI-supported group had a detection rate of 6.7 cancers per 1,000 women, versus 5.7 per 1,000 with standard double reading, while recall rates — how often women were called back — were similar (37.4 versus 38.3 per 1,000).

That is a meaningful signal from real practice. But PRAIM was observational: exams were not randomly assigned to one approach or the other, so other differences between the groups could have contributed. And neither study can yet tell us the ultimate outcome that screening exists for — fewer deaths from breast cancer — which takes many more years to measure.

Hidden costs: false alarms, overdiagnosis and deskilling

Three concerns deserve attention as AI tools spread beyond these carefully studied programs.

1. Different settings, different error rates. A tool tuned in one population may raise more false alarms, or miss more cancers, in another with different equipment, ages or breast density. Local monitoring matters.
2. More detection could mean more overdiagnosis. Extra cancers found at screening are good news only if they are ones that would have caused harm. MASAI’s finding of fewer aggressive interval cancers is encouraging on this point; it is not yet the final word.
3. Human skills may fade. A study in The Lancet Gastroenterology & Hepatology, covering four colonoscopy centres in Poland, found that experienced doctors’ adenoma detection rate — the share of colonoscopies in which they find precancerous growths — fell from 28.4% to 22.4% in procedures done without AI after they had grown used to working with it. The study was observational and other factors may have contributed, but it is a warning that humans in the loop need to stay sharp.

What regulators have cleared — and what that means

The U.S. Food and Drug Administration maintains a public list of AI-enabled medical devices authorized for marketing, and radiology accounts for the largest share of it. Authorization means a device has met FDA’s requirements for a specific intended use. It does not, by itself, show that the device improves long-term outcomes such as survival, and the FDA notes the list is not comprehensive.

AI is also being studied for lung CT, skin lesions, pathology slides and more. Much of that work is still retrospective — testing software on old images with known outcomes — rather than prospective trials in live screening programs. Retrospective accuracy is a necessary first step, not proof of benefit.

What to watch

• Longer follow-up from MASAI and other trials, including effects on advanced cancers and, eventually, deaths.
• Results from programs with different machines, populations and AI products.
• Whether recall and biopsy rates stay flat as AI use scales up.
• How programs keep radiologists’ independent skills intact.

For a companion look at how AI is changing the scans themselves, read The MRI Reinvention. For why earlier detection matters so much in cancer, see Is There a Cure for Cancer? And before sharing the next “AI beats doctors” headline, run it through our Breakthrough Reality Check.