The MASAI Trial — Testing AI-Supported Breast Screening in a Randomised Controlled Trial

The MASAI trial (Mammography Screening with Artificial Intelligence) is a Swedish study that compared AI-supported mammography screening with the standard method in a randomised controlled trial. Led by Kristina Lång of Lund University, it announced on 2 August 2023 that its interim safety analysis had been published in The Lancet Oncology: the AI arm detected 20% more cancers (a difference that was not statistically significant) and cut radiologists' screen-reading workload by 44%. A February 2025 analysis raised the increase in detection to 29%, and the final analysis in January 2026 showed that the rate of interval cancers, the primary endpoint, was non-inferior to standard reading. Lund University describes MASAI as the first randomised controlled trial evaluating the effect of AI-supported screening.

Plenty of studies had shown AI reading mammograms before, but most were retrospective, using archived images. The Lancet Oncology paper opens by noting that retrospective studies had shown promising results but that, to the authors' knowledge, no randomised trial had yet been done.

How the trial worked

Around one million women are invited to mammography screening in Sweden each year, and each examination is read independently by two breast radiologists, known as double reading. According to Lund University, there is a workforce shortage of breast radiologists in Sweden and elsewhere that can put screening services at risk.

The trial ran at four screening sites in southwest Sweden: Malmö, Lund, Landskrona and Trelleborg. Women who received the screening invitation and did not opt out were randomised after image acquisition by the picture archiving system.

Participants and radiographers were masked to the allocation; the radiologists reading the images were not. The key point is that the AI did not make diagnoses on its own. It redistributed radiologists' numbers and attention.

Three analyses and their results

In 2023 the extra detection, 41 more cancers, did not reach statistical significance. That analysis was a safety check: the AI arm's detection rate cleared the pre-set lower limit of 3 per 1,000, so the trial was allowed to continue.

By 2025 the difference was clear. The extra cancers were mainly small, lymph-node-negative invasive cancers, and about half of the extra in situ cancers were high-grade, with no increase in the lowest grade. The positive predictive value of recall, the share of women called back who actually had cancer, was also higher with AI.

The final analysis's primary endpoint was interval cancer: cancer diagnosed after a normal screen and before the next scheduled round. Lång explained that interval cancers are typically more aggressive and that their frequency is a central measure of how effective a screening test is. The AI arm had 82 interval cancers and the control arm 93. The trial was designed with a 20% non-inferiority margin and showed that AI support was not worse than standard reading; but the confidence interval around the ratio of 0.88 crosses 1, so a reduction as such was not statistically established. The paper reports descriptively that the AI arm had fewer interval cancers that were invasive (75 vs 89), T2 or larger (38 vs 48) or non-luminal A (43 vs 59).

Limitations the researchers named

In Lund University's 2023 announcement Lång said: "The study was conducted on a single site in a Swedish setting. We need to see whether these promising results hold up under other conditions, for example with other radiologists or other AI algorithms." She added that other ways of using AI in screening should also be investigated prospectively.

The time saving was likewise an estimate. Assuming a radiologist reads 50 examinations an hour, Lund University calculated that reading the roughly 40,000 examinations in the AI arm took about five months less of a radiologist's time.

The researchers' interests are declared in the papers. Lång has been an advisory board member for Siemens Healthineers, and one co-author heads BreastScreen Norway at the Cancer Registry of Norway, which has a research agreement with ScreenPoint Medical, the maker of Transpara.

Finding more in situ cancers also raises the question of overdiagnosis, detecting lesions that might never have needed treatment. The 2025 paper showed that about half of the extra in situ cancers were high-grade and that the lowest grade did not increase. The extent of overdiagnosis itself was not an endpoint of the trial.

Who it affects

The people most directly affected are radiologists. Under the MASAI approach, many examinations are read by one radiologist instead of two, which Lund University framed as a way to relieve a scarce specialty. For the women being screened, Lund University says nothing noticeable changes: the examination proceeds as usual and the AI is used afterwards, when the images are read.

In January 2026 Lång said that many regions in Sweden had already started to implement AI-supported screening, with more expected to join, and that because the method is simple it would probably have an international impact. On 4 February 2025 the UK Department of Health and Social Care announced EDITH, a trial of AI in breast screening involving nearly 700,000 women at 30 sites, backed by £11 million through the National Institute for Health and Care Research. The department said that while two specialists are currently needed per mammogram, the technology enables one to do the same work safely and efficiently, and that a successful trial could free up hundreds of radiologists and other specialists; that is the department's claim.

The 2018 editorial on IDx-DR had warned that accuracy studies alone cannot show whether patients benefit. MASAI built the AI into the real screening workflow, randomised the comparison, and made interval cancer, a measure of screening effectiveness, its primary endpoint. That is the reverse of the order seen with the Epic Sepsis Model, which was externally validated only after wide use. For the wider period, see the AI timeline and the AI and society timeline.