What changed
In Sweden's MASAI trial, mammograms were either read in the usual way by two radiologists or triaged by an AI system. Lower-risk AI cases received one radiologist read; the highest-risk cases received two reads, with AI marks available as decision support.
The AI-supported group reached 80.5% sensitivity, compared with 73.8% for standard double reading. Specificity was 98.5% in both groups. The rate of cancers diagnosed between screening rounds was non-inferior: 1.55 per 1,000 participants with AI support versus 1.76 per 1,000 with standard reading.
What this could change for you
For a person getting a routine mammogram, the useful change is not an AI diagnosis. It is a better second set of eyes inside the screening workflow, with radiologists still making the decision.
The result also matters for access. Earlier MASAI analyses found that this setup reduced screen-reading workload while detecting more cancers, which could help screening programs use scarce specialist time more effectively.
What it does not prove
This was one organized Swedish screening program using a particular commercial system. Results may not transfer unchanged to every country, clinic, mammography machine, or population.
The trial measured screening performance, not whether AI support reduces breast-cancer deaths. It does not support uploading a mammogram to a general-purpose AI tool or skipping a radiologist.
The bottom line
This is unusually strong evidence for medical AI: a large randomized trial in a real screening program, with better sensitivity and no loss of specificity. The change is ready for health systems to evaluate—not for patients to use on their own.
Primary research
Interval cancer, sensitivity, and specificity comparing AI-supported mammography screening with standard double reading without AI in the MASAI study
The Lancet · 2026 · DOI 10.1016/S0140-6736(25)02464-X
