TrialsgrambetaRCTs are beautiful

MASAI

Mammography Screening with Artificial Intelligence trial

Patient / Population Intervention / Exposure Comparison Outcome

In 105,934 women in population-based mammography screening, AI-supported screen reading was non-inferior to standard double reading without AI for interval cancer rate, with higher cancer detection and a 44.2% lower screen-reading workload.

N
105,934patients
Design
Population-based single-blinded non-inferiority RCT, Sweden
Endpoint
Interval cancer rate (20% non-inferiority margin)
Relevance
1Practice-defining — the trial the guideline rests on.
ResultInterval cancers 1.55 vs 1.76 per 1,000 (ratio 0.88, 95% CI 0.65–1.18; non-inferior); sensitivity 80.5% vs 73.8% (P=0.031), specificity 98.5% in both. Cancer detection 6.4 vs 5.0 per 1,000 (ratio 1.29, 95% CI 1.09–1.51; P=0.0021); screen-reading workload −44.2%. Interim safety analysis in 80,033 women: detection 6.1 vs 5.1 per 1,000 (ratio 1.2, 95% CI 1.0–1.5), workload −44.3%.
Lång K, et al. Lancet Oncol. 2023;24(8):936-944; Hernström V, et al. Lancet Digit Health. 2025;7(3):e175-e183; Gommers J, et al. Lancet. 2026;407(10527):505-514. 10.1016/S1470-2045(23)00298-X
Discussion & critique

The first randomised trial of AI in mammography screening and the anchor of the cluster: AI triage nearly halved radiologist reading with no loss of safety, more (mostly small, node-negative) cancers detected, and a non-inferior interval cancer rate at two years. One AI system (Transpara) at four sites of one Swedish programme with double reading as comparator, so the result does not transfer directly to single-reader programmes; whether earlier detection lowers breast cancer mortality is beyond the trial. PRAIM and ScreenTrustCAD are the non-randomised companions.