AI & Medicine6 min read
AI and Cancer: What Artificial Intelligence Really Delivers Today
Large studies from Sweden and Germany show where AI improves cancer screening, and why doctors still make the decisions.

“AI cures cancer” is how some headlines read. It is not true. What is documented is more sober: in early detection, AI can work as an additional, tireless reader next to doctors, and that benefit can be measured in large studies. For companies this is interesting less for the medicine than for the method used to prove the benefit.
Breast cancer screening: the best evidence so far
The strongest evidence so far comes from Sweden. In the MASAI trial (Mammography Screening with Artificial Intelligence), women in breast cancer screening were randomly assigned to one of two groups: AI-supported reading, or the usual double reading by two radiologists without AI. The safety analysis published in 2023 covered 80,033 women aged 40 to 80. The AI group had 6.1 cancers detected per 1,000 screened, the control group 5.1; the authors judged this a similar detection rate [1]. The difference in effort was clear:
44.3 %
There was no warning sign on the side effects of screening either: 2.2 percent of women with AI and 2.0 percent without were invited for further tests, and the share of false positives was 1.5 percent in both groups [1]. That matters, because a method that finds more tumours but needlessly worries many healthy women would be no progress. The authors themselves read the results as evidence that the approach is safe, not as a free pass: the question about interval cancers was still open at that point [1].
In January 2026 the final results followed, with 105,934 randomly assigned women. The key question was now a different one: does AI-supported reading end up missing more tumours? The measure was so-called interval cancers, which are tumours that appear between two screening appointments even though the earlier mammogram looked normal. Their rate was 1.55 per 1,000 women with AI and 1.76 without [2].
1.55 vs. 1.76
Sensitivity, the share of existing tumours that are found, was 80.5 percent with AI against 73.8 percent, while specificity was 98.5 percent in both groups, so there were no more false alarms [2]. One caveat matters here: the trial was designed to show that the AI approach is not worse, and it did. The drop in interval cancers itself was not statistically established (ratio 0.88, confidence interval 0.65 to 1.18) [2]. The honest core message is therefore: at least equal quality with much less work, not proof of “fewer deaths”.
Descriptively, there were also fewer invasive interval cancers with AI (75 against 89 cases) [2]. These numbers should be read as a pointer, not as a proven effect: the trial was not designed for that, and for in situ tumours, the early stages, the sensitivity advantage did not show up [2].

And in Germany?
An answer from everyday practice comes from the PRAIM study, published in Nature Medicine in 2025. It evaluated the organised mammography screening programme at twelve sites in Germany: 463,094 women aged 50 to 69, of whom 260,739 were screened with AI support, read by 119 radiologists [3]. The detection rate was 6.7 per 1,000 women with AI and 5.7 without.
+17.6 %
The recall rate, meaning how many women are called back for further tests, did not rise: 37.4 against 38.3 per 1,000. And of the recalls, more were confirmed with AI (17.9 against 14.9 percent) [3]. The catch: PRAIM was not a randomised trial. Radiologists chose voluntarily whether to use the AI [3]. Such observational data are less robust than a comparison with random assignment, but they show that the effect also arrives in the daily routine of large screening centres, not only under study conditions.
Pathology: a second look at the tissue
There is also authorised software in the lab. In 2021 the US regulator FDA granted de novo authorisation to “Paige Prostate”, which according to its maker was the first AI-based pathology product authorised that way [4][5]. The software marks areas in scanned prostate tissue slides that look suspicious for cancer. The FDA defines its role clearly: an adjunctive, computer-assisted method whose output should not be used as the primary diagnosis [4].
It is worth seeing how soberly the agency describes the benefit. In the pivotal study, detection improved in 7.3 percent of tissue samples that contained cancer. At the same time the FDA notes that the benefit per patient in practice is likely substantially lower, because a biopsy takes several samples and many patients have cancer in more than one of them [4].

Where the limits are
- The authorisation covers only the defined use: Paige Prostate is intended for prostate needle biopsies, specific scanners and the matching viewing software [4].
- What was measured is detection and interval cancers, not mortality [2]. These studies do not say whether AI screening saves lives in the long run.
- The human stays in the process: in MASAI the AI sorted exams to single or double reading by radiologists [2], in PRAIM radiologists read with AI support [3], and the FDA requires Paige Prostate to be used alongside the full standard-of-care review [4].
- Errors remain possible: the FDA names false positive and false negative results explicitly as a risk that is limited only by requirements and trained users [4].
The yardstick: benefit you can verify
What sets these studies apart is less the technology than the discipline. MASAI defined in advance what “not worse” means, a margin of 20 percent, and registered the trial publicly [2]. There was a comparison group, a fixed target measure and a regulator that draws the field of use narrowly. The AI is not sold as a replacement but as a second opinion whose contribution can be counted.
This article is not medical advice.
Sources (5)
- [1]Artificial intelligence-supported screen reading versus standard double reading in the Mammography Screening with Artificial Intelligence trial (MASAI): a clinical safety analysis (Lång et al.)The Lancet Oncology 24(8), 936–944 (entry in the Lund University Research Portal) · August 1, 2023
- [2]Interval cancer, sensitivity, and specificity comparing AI-supported mammography screening with standard double reading without AI in the MASAI study (Gommers et al.)The Lancet 407(10527), 505–514 (entry in the Lund University Research Portal) · January 31, 2026
- [3]Nationwide real-world implementation of AI for cancer detection in population-based mammography screening (PRAIM)Nature Medicine 31(3), 917–924 (entry of the University of Lübeck) · January 7, 2025
- [4]Evaluation of Automatic Class III Designation for Paige Prostate – De Novo Decision Summary DEN200080U.S. Food and Drug Administration
- [5]Paige Receives First Ever FDA Approval for AI Product in Digital Pathology (manufacturer press release)Paige / BioSpace · September 22, 2021
How much of this is in your day-to-day?
In a free initial consultation we look together at what AI can already do in your business today – and where it has no place.
Book an initial consultation

