About this
The MASAI trial, led by Kristina Lång at Lund University, randomly split about 106,000 women at four Swedish screening sites into AI-supported screening or the usual reading by two doctors. This is an independent explainer. The researchers didn’t make or endorse it.
How the picture works
- Cancers found: 5.0 vs 6.4 per 1,000 women (ratio 1.29, 95% CI 1.09–1.51), shown per 10,000.
- Doctor work: 109,692 vs 61,248 readings, 44.2% fewer. Per woman that’s about 2.1 vs 1.2 readings (our calc). Bright dots are women whose scan got a second doctor. Without AI that’s nearly everyone. With AI it’s about 15% (our approximation from the totals).
- False alarms: callbacks that didn’t find cancer, recalls minus cancers found (1,027 − 262 vs 1,110 − 338). That works out to about 145 vs 146 per 10,000 (our calc). The paper’s false-positive ratio is 1.01 (0.91–1.11).
- Which dots light up is random. Only the counts are real. The dots then gather into rows (10 per row for cancers, 20 for false alarms) so the totals are easy to compare. The dashed line marks the usual-way total.
Still unknown
Whether this saves lives (the trial counted cancers found, not deaths), how much it leads to treating cancers that would never have caused harm, and how well one AI tool from Sweden, where two doctors normally read each scan, carries over elsewhere. Cancers that showed up between screenings were slightly lower with AI (1.55 vs 1.76 per 1,000), but that difference wasn’t statistically clear.
Sources
- Hernström V, et al. Lancet Digital Health 7, e175–e183 (2025). Paper · CC BY 4.0
- Gommers J, et al. The Lancet 407, 505–514 (2026). Paper
AI: Transpara 1.7.0 (ScreenPoint Medical). Funded by the Swedish Cancer Society and Swedish public research funds. Some authors report industry ties. See the papers’ declarations.