MASAI: AI-supported mammography screening finds more cancers with half the radiologist workload
In the first randomised trial of AI in population mammography, AI-supported reading detected 20% more cancers than standard double reading without increasing false positives, and cut screen-reading work by 44%.
MASAI randomised 80,033 women aged 40-74 attending screening in Sweden to AI-supported screening (AI risk score used to triage to single or double reading and to flag suspicious findings) or standard double reading. This pre-specified safety analysis compared cancer detection, recall and false positives.
The cancer detection rate was 6.1 per 1,000 with AI support versus 5.1 per 1,000 with standard reading (ratio 1.2), with recall rates of 2.2% and 2.0% and identical 1.5% false-positive rates. Screen-reading workload fell by 44.3%. The 2025 report on the full cohort (over 105,000 women) confirmed a 29% increase in detection, concentrated in invasive and small cancers.
This was the first randomised evidence that AI can safely replace one of two human readers in screening.
- Cancer detection rate 6.1 vs 5.1 per 1,000 screened (ratio 1.2, 95% CI 1.0-1.5; 244 vs 203 cancers)
- Recall rate 2.2% vs 2.0%; false-positive rate 1.5% in both arms
- Screen-reading workload reduced 44.3%
- Final analysis (Lancet Digital Health 2025, 105,934 women): detection 6.4 vs 5.0 per 1,000, a 29% increase, with a 44% workload reduction
AI can take over one reader's work in double-reading screening programmes while finding more cancers. Whether the extra cancers found are ones that would have harmed women, and whether interval cancers fall, is the question the trial's primary endpoint will answer.
- Interval cancer rate, the primary endpoint, has not yet been reported; more detection could be overdiagnosis
- Single Swedish site with a double-reading culture; generalisation to single-reader systems (as in the US) is uncertain
- One commercial AI system; results do not automatically transfer to others
- Radiologists were aware of AI output, so the trial cannot separate AI accuracy from its effect on human reading