What is the problem?
In many countries, each screening mammogram is read by two radiologists (double reading). It works well but takes a lot of specialist time. Earlier studies had indicated that AI could find more cancers and cut the reading workload.
One important question remained: does AI change the number of interval cancers, meaning those diagnosed between screens or within 2 years of the last screen without having been seen at screening?
What did the researchers do?
MASAI is a Swedish population-based randomised trial. Between April 2021 and December 2022, 105,934 women were assigned in equal numbers to AI-supported screening or standard double reading without AI. 19 were excluded from the analysis.
In the AI group, the program sorted exams to single or double reading by radiologists, and acted as detection support. The main outcome was the interval cancer rate, with a 20% non-inferiority margin, meaning the aim was to show AI was not clearly worse.
What were the results?
The interval cancer rate was 1.55 per 1,000 participants with AI against 1.76 with double reading, a ratio of 0.88: non-inferiority was established. Interval cancers with unfavourable features were also fewer, for example 75 invasive cancers against 89.
Sensitivity (the share of cancers found) was higher with AI: 80.5% against 73.8%. Specificity (the share of women without cancer correctly cleared) was identical at 98.5%. In plain terms, among women who had a cancer, AI-supported screening caught more of them, without multiplying false alarms.
Key figures
- 105,934women randomly assigned to AI-supported screening or double reading without AI
- 1.55 vs 1.76interval cancers per 1,000 participants, with AI against without AI
- 80.5% vs 73.8%sensitivity (cancers found), with AI against without AI
What does this change for patients?
The results suggest AI-supported screening can find more cancers without raising false alarms, while easing the reading workload. The authors say it may be considered for clinical practice.
For a woman invited to screening, nothing changes today unless her screening programme decides so: keep following your country's recommendations.
What are the limits?
The main aim was non-inferiority: the trial does not show that AI clearly lowers the number of interval cancers (the confidence interval of the ratio, 0.65 to 1.18, includes 1). The sensitivity advantage did not apply to in-situ (non-invasive) cancers.
The sensitivity difference is only moderately significant (p=0.031), and other results from the trial have been reported separately. The trial took place in one country, with a particular screening programme and a particular AI. The results need to be confirmed elsewhere before being applied more widely.
Breast screening: keeping track with Mima
The MASAI trial suggests AI-supported screening can be considered. It depends on the country and the centre, and not everyone has access.
With Mima, you can note your mammograms, their dates and results, and prepare your question: "Does my centre use AI to help read the scans, and what does it change for me?" Ask it at your next appointment.
Frequently asked questions
Can AI replace the radiologist in breast cancer screening?
No. In MASAI, AI sorted exams to single or double reading by radiologists and helped them detect cancers. It did not replace doctors.
What is an interval cancer?
It is a breast cancer diagnosed between two screens, or within 2 years after the last scheduled screen, without having been detected at that screen.
Is screening with AI safer?
It proved at least as safe: an interval cancer rate of 1.55 against 1.76 per 1,000, with higher sensitivity and identical specificity.
What is a non-inferiority trial?
It is a trial designed to show a new method is not clearly worse than the reference method, here with a 20% margin.
The original study
Gommers J, Hernström V, Josefsson V, Lång K, et al. “Interval cancer, sensitivity, and specificity comparing AI-supported mammography screening with standard double reading without AI in the MASAI study: a randomised, controlled, non-inferiority, single-blinded, population-based, screening-accuracy trial.” The Lancet, January 2026. DOI : 10.1016/S0140-6736(25)02464-X
Read the full study in The LancetThis article summarises a scientific study and is not medical advice.