AI-Supported Mammogram Reading Caught More Cancers Without More False Alarms

CATEGORY: WOMEN’S HEALTH

Two radiologists read every screening mammogram in Sweden’s national program, each one working independently. A trial of 105,934 women handed part of that job to artificial intelligence, and detection went up rather than down: 80.5% of breast cancers caught, against 73.8% with the standard two-reader approach, and no increase in false alarms.¹ It is the first trial of its kind to report complete results.

An AI Read First, Then Sent Most Women to a Single Radiologist

Double reading is a safeguard, and an expensive one. MASAI, as the trial was called, tested a rearrangement. An AI system examined each screening mammogram before any human did, marked suspicious areas for the radiologists, and routed the clearly low-risk majority to one reader instead of two. Screen-reading workload fell 44%, an earlier report from the same trial found.² What nobody yet knew was whether the cancers were still being found.

Sweden’s Cancer Registries Counted What Screening Missed

Women arriving for routine screening at four Swedish sites between April 2021 and December 2022 were split half and half at random, AI-supported reading on one side, standard double reading on the other. Then came the slow part. For roughly two years, until each woman’s next scheduled visit, Sweden’s cancer registries recorded every breast cancer diagnosed in either group. That included interval cancers, the ones that surface between visits after a normal result. Counting those is what turns scans flagged into cancers actually caught.

Interval Cancers Ran 1.55 per 1,000 Women Versus 1.76

Screening in the AI-supported group caught 80.5% of the cancers that emerged, versus 73.8%, and the advantage held across age groups and breast densities. False alarms stayed flat. In both groups about 98.5% of women without cancer were correctly left unrecalled. Interval cancers came in at 1.55 per 1,000 women with AI support against 1.76 without, a gap small enough to be chance, though fewer of the AI group’s missed cancers were invasive, large or aggressive subtypes.¹ Invasive interval cancers numbered 75 in the AI group and 89 in the other.¹

Every Scan Still Had a Human Reader

The AI assisted radiologists; it did not replace them. At least one person read every scan. A callback is not a diagnosis either, and most recalls turn out to be benign findings such as cysts or overlapping tissue. The trial measured cancer detection, not deaths, and whether earlier detection means fewer breast cancer deaths would take longer follow-up to establish. Sweden runs an organized screening program, so countries that already use single reading, the United States among them, may see something different. Germany’s national program pointed the same way: among 463,094 women, radiologists using AI support detected about 18% more breast cancers, 6.7 per 1,000 screened versus 5.7, with no rise in callbacks for further tests.³ That was routine practice rather than a randomized comparison, which makes it an association and nothing firmer.

Key Takeaways

• AI-supported reading caught 80.5% of breast cancers in the trial, against 73.8% for two radiologists reading every scan, and false alarms did not rise.
• Cancers surfacing between screening visits were no more common with AI support, and fewer were invasive or aggressive, though that difference was small enough to be due to chance.
• Radiologists were assisted rather than replaced, and an earlier report from the same trial put the screen-reading workload saving at 44%.
• Detection is not mortality. Whether AI-supported screening leads to fewer deaths from breast cancer is beyond what this trial can answer.

References

1. Gommers J, Hernström V, Josefsson V, et al. Interval cancer, sensitivity, and specificity comparing AI-supported mammography screening with standard double reading without AI in the MASAI study: a randomised, controlled, non-inferiority, single-blinded, population-based, screening-accuracy trial. The Lancet. 2026;407(10527):505–514. link
2. Lång K, Josefsson V, Larsson AM, et al. Artificial intelligence-supported screen reading versus standard double reading in the Mammography Screening with Artificial Intelligence trial (MASAI): a clinical safety analysis of a randomised, controlled, non-inferiority, single-blinded, screening accuracy study. The Lancet Oncology. 2023;24(8):936–944. link
3. Eisemann N, Bunk S, Mukama T, et al. Nationwide real-world implementation of AI for cancer detection in population-based mammography screening. Nature Medicine. 2025;31(3):917–924. link.