Randomised trial of AI-supported mammography screening in 105,934 women (MASAI)
In a Swedish screening trial that randomised 105,934 women, AI triage and detection support found 29% more cancers than standard double reading, cut radiologists' reading workload by 44%, and met its primary endpoint of a non-inferior interval cancer rate, 1.55 against 1.76 per 1,000, with sensitivity of 80.5% against 73.8% at the same 98.5% specificity.
- Model
- Transpara 1.7.0 (ScreenPoint Medical)
- Field
- Medicine
- Date
- 2026-01-30
- Human collaborators
- Kristina Lång, Veronica Hernström, Jessie Gommers
Sources
Original work
Announcement
Media coverage
What was found
Women screened at four sites in southwest Sweden between April 2021 and December 2022 were randomised 1:1. In the AI arm, Transpara gave each examination a malignancy risk score from 1 to 10: scores 1 to 9 went to a single radiologist and score 10 to two, and radiologists saw the scores and, for high scores, the system's marks on suspicious areas. The control arm was the Swedish standard of independent double reading without AI. Results came in three papers. A 2023 safety analysis of the first 80,033 women (Lancet Oncology) found the AI arm safe, with a 44.3% lower reading workload. A 2025 analysis of all participants (Lancet Digital Health) found 338 cancers against 262, a detection rate of 6.4 against 5.0 per 1,000 (ratio 1.29), mostly small node-negative invasive cancers, with no significant rise in recalls or false positives and 61,248 against 109,692 screen readings. The 2026 primary-endpoint paper (The Lancet) reports interval cancers, those surfacing between screening rounds: a proportion ratio of 0.88 (95% CI 0.65 to 1.18) within a 20% non-inferiority margin, with descriptively fewer invasive (75 against 89), large (38 against 48) and non-luminal A (43 against 59) interval cancers.
Novelty check
Europe PMC, (artificial intelligence OR deep learning OR AI) AND mammograph* in titles, with randomised in the title or as publication type, to July 2023: no earlier randomised trial; the single hit is a retrospective model study. The 2023 safety paper states that no randomised trial had been done. ScreenTrustCAD in Stockholm evaluated AI as an independent reader prospectively but with a paired, not randomised, design.
Caveats and known objections
One country whose standard of care is double reading, four sites and one AI product and version, so the gain over single-reader programmes such as the US one is not measured here. The trial was designed for non-inferiority, and the interval cancer reduction is not statistically significant: the confidence interval runs from a 35% reduction to an 18% increase, so coverage reporting 12% fewer interval cancers overstates what the trial showed. The breakdown by cancer type is descriptive. Radiologists were not masked to allocation, since they saw the AI scores. The trial was publicly funded and run by academic investigators; the AI is ScreenPoint Medical's commercial product. Autonomy is ai-assisted because radiologists read every examination and made every recall decision.
Trial registration
ClinicalTrials.govMammography Screening with Artificial Intelligence (MASAI)
NCT04838756
That a study was registered, with the outcomes it declared and the date it declared them, so a published endpoint can be checked against the one planned. It does not certify the results, their analysis, or any clinical benefit; registry records are submitted by the sponsor and are not peer reviewed.
Regulatory authorization
FDA AI-enabled devicesTranspara 1.7.0
K210404
That the FDA authorized the device for marketing in the US for a stated indication, on the evidence the manufacturer submitted. Most AI devices reach the market by 510(k), which requires substantial equivalence to an earlier device rather than new clinical evidence, so an authorization is not evidence of clinical benefit.
The version used in the trial. Later versions hold their own clearances, the most recent in the FDA list being K241831 for Transpara 2.1.0.
Nobody outside the lab has checked this yet.
Reading the primary source closely enough to say whether it supports the claim counts as a check, and you are credited on the entry.
Or on GitHub: submit a check challenge the grade send a correction or send a pull request
Flag this for triage
Signals order the review queue and nothing else. They are never published, and they never move a grade: that takes a citation.
Entry history (1 event)
- AddedEntered the registry graded Peer reviewed and AI-assisted.
Entries are never deleted. A grade that does not hold up is downgraded on the record, with the reason beside it.
Graded peer reviewed for verification and ai-assisted for autonomy. What these mean.
Cite this entry
whataifound.org. (2026). Randomised trial of AI-supported mammography screening in 105,934 women (MASAI). whataifound.org: A Registry of AI Scientific and Mathematical Discoveries. https://whataifound.org/finding/2026-01-30-masai-ai-mammographyBibTeX
@misc{whataifound-lunduniversity-2026-mammography,
title = {Randomised trial of AI-supported mammography screening in 105,934 women (MASAI)},
author = {{whataifound.org}},
year = {2026},
howpublished = {whataifound.org: A Registry of AI Scientific and Mathematical Discoveries},
note = {Result by Lund University. Verification: Peer reviewed. Autonomy: AI-assisted.},
url = {https://whataifound.org/finding/2026-01-30-masai-ai-mammography}
}