AI breast screening found more cancers and cut radiologists' reading work by 44%
In Sweden's MASAI trial, AI-supported screening caught 80.5% of cancers, against 73.8% for two radiologists, with no rise in false alarms.
Sweden’s MASAI trial is the strongest test so far of AI in breast screening. Between April 2021 and December 2022 it assigned 105,934 women at four sites, at random, to one of two routes. Either an AI sorted their mammogram and marked suspicious areas for the radiologist, or two radiologists read every image with no AI, which is the European standard.
The AI route caught more cancers: 80.5% of them against 73.8% (p=0.031). False alarms did not go up, and were identical in both groups at 98.5% of healthy women correctly cleared. Detection figures published a year earlier worked out at 6.4 cancers found per 1,000 women with AI against 5.0 without, with more of the small invasive cancers that matter most. Radiologists did 44% fewer readings in the AI group, because most images went to one reader instead of two.
The trial’s main question was about cancers that turn up in the gap between screening rounds, meaning screening missed them. There were 82 of these with AI and 93 without. That is a reduction in the right direction, but it is too small to be sure it is real (p=0.41). What the trial did establish is that AI support is no worse on this measure.
A German programme, PRAIM, points the same way among 463,094 women, but radiologists there chose whether to use the AI. A 2026 review that pooled the stronger trials found a smaller and uncertain benefit, and concluded AI belongs in screening as a second reader, not as a replacement for the radiologist.
What it means
For women invited to breast screening in a programme using a system like this, the evidence says more cancers get found without more false alarms, and the radiologist still signs off. For health services, the 44% cut in reading work matters most: it is a way to screen more people with the radiologists they already have. What nobody has shown yet is that this means fewer women die of breast cancer.
Caveats
- MASAI tested one AI system (Transpara) at four Swedish sites with one type of equipment. The 2026 review found accuracy varies so much between programmes that these results cannot be assumed to hold elsewhere without local checking.
- The trial showed AI was no worse on missed cancers, not that it reduced them. The counts were 82 against 93.
- No trial has yet shown that AI-supported screening means fewer women die of breast cancer.
- PRAIM only observed what radiologists chose to do, which can flatter the AI. The review notes the bigger gains come from such designs, the pattern you would expect if something else were at work.
- Finding more cancers is not automatically good: some would never have caused harm. MASAI reports which kinds it found, which is why the extra invasive and higher-grade cancers matter more than the raw count.