What is AI mammography screening?
AI mammography screening means using computer software to help radiologists read screening mammograms. The software looks at each set of breast X-ray images, gives it a risk score and can mark areas that look suspicious. The radiologist still reads the images and decides whether a woman should be called back for more tests.
The largest test of this idea so far is the MASAI trial in Sweden. In MASAI, about 106,000 women having routine screening were randomly split into two groups. One group had AI-supported reading. The other had the usual Swedish method, where two radiologists read every mammogram, called double reading. The mammogram itself was the same in both groups; only the reading afterwards was different.
This page explains how the trial was set up, what it found on cancers detected, cancers that appeared between screening rounds and radiologist workload, what other studies show, the limits of the evidence and what it may mean for women in India.
Key facts about the MASAI trial of AI mammography screening
The main published results, in brief:
- Design: randomised, controlled, non-inferiority trial at four screening sites in southwest Sweden
- Enrolment: 105,934 women randomised between April 2021 and December 2022; 19 were later excluded
- AI system: Transpara version 1.7.0 (ScreenPoint Medical), used to sort exams to one or two readers and to mark suspicious areas
- Cancers found at screening (Lancet Digital Health, 2025): 338 with AI support vs 262 with standard double reading (6.4 vs 5.0 per 1,000 women)
- Screen-reading workload: 44.2% lower with AI support (61,248 vs 109,692 readings)
- Interval cancers (The Lancet, 2026): 1.55 vs 1.76 per 1,000 women (82 vs 93 cancers); AI support met the trial's non-inferiority goal
- Sensitivity 80.5% vs 73.8%, with the same specificity of 98.5% in both groups
Key takeaways: AI mammography screening and the MASAI trial
- In MASAI, AI-supported reading found more breast cancers at screening than two radiologists reading without AI, without more false positives.
- Radiologists did about 44% fewer screen readings when AI sorted the low-risk mammograms to a single reader.
- Cancers that appeared between screening rounds were slightly fewer with AI (82 vs 93); this was enough to show AI was not worse, but not enough to prove it was better.
- Every mammogram was still read by at least one radiologist. The authors say the study does not support replacing health professionals with AI.
- The trial was done in Sweden's organised screening programme with one AI product, so results may not carry over directly to India.
How did AI support mammogram reading in the MASAI trial?
In standard Swedish screening, two radiologists read every mammogram separately. In the AI-supported group, the Transpara software first gave each exam a malignancy risk score on a 10-level scale. Exams with scores of 1 to 9 were read by one radiologist. Exams with the highest score, 10, were read by two radiologists. This is called triage.
The AI also acted as a second pair of eyes. Radiologists could see the risk score for every exam, and for exams scoring 8 to 10 the software marked the areas it found suspicious. The radiologists then decided whether to call the woman back for further tests.
For the woman herself, nothing changed. Lund University, where the trial was led, notes that the mammogram was taken in the usual way and the AI was used only afterwards, when the images were read.
How was the MASAI trial designed and who took part?
MASAI was a randomised controlled trial, the strongest type of study for comparing two approaches. Women attending routine mammography screening at four sites in southwest Sweden (Malmö, Lund, Landskrona and Trelleborg) were randomly assigned, in equal numbers, to AI-supported reading or standard double reading. Women and radiographers did not know which group a woman was in, but the radiologists did, because they could see whether AI was in use. In Sweden, women aged 40 to 74 are regularly invited to screening.
Between April 2021 and December 2022, 105,934 women were randomised. After 19 were excluded, 53,043 were analysed in the AI group and 52,872 in the standard group.
The trial was designed as a non-inferiority study. Its main question was whether AI-supported screening was no worse than standard double reading on interval cancers, using a pre-set margin of 20%. An early safety analysis of the first 80,033 women was published in 2023, the screening results for all women in 2025, and the interval cancer results in 2026.
Did AI mammography screening find more breast cancers?
Yes, in this trial. In the full group of women, published in The Lancet Digital Health in 2025, AI-supported reading found 338 cancers compared with 262 with standard double reading. That is 6.4 compared with 5.0 cancers per 1,000 women screened, about 29% more, and the difference was statistically significant.
The extra cancers were of two kinds. There were more invasive cancers (270 vs 217), and the authors report these were mainly small tumours that had not spread to lymph nodes. There were also more in situ cancers (68 vs 45), meaning abnormal cells that have not grown beyond the milk ducts; about half of the extra in situ cancers were high grade.
The earlier 2023 safety analysis of 80,033 women showed the same direction: 244 cancers with AI support and 203 without (6.1 vs 5.1 per 1,000). Finding more cancers at screening is not the same as saving more lives; that needs longer follow-up.
Did AI reduce interval cancers in the MASAI trial?
Interval cancers are breast cancers diagnosed between screening rounds, after a mammogram was reported as normal. They matter because they include cancers that screening missed, and they were the trial's main measure.
The results were published in The Lancet in January 2026, after every woman had two years of follow-up. The interval cancer rate was 1.55 per 1,000 women with AI support and 1.76 per 1,000 with standard double reading (82 vs 93 cancers). That is about 12% fewer, but the confidence interval for the difference was wide (rate ratio 0.88, 95% CI 0.65 to 1.18), so the trial shows AI support was not worse, not that it was proven better.
Sensitivity, the share of all cancers found at screening, was higher with AI (80.5% vs 73.8%), and specificity was the same (98.5%). The gain was seen for invasive cancers but not in situ cancers. There were also fewer invasive (75 vs 89), larger T2 or above (38 vs 48) and non-luminal A (43 vs 59) interval cancers with AI, but the authors describe these comparisons as descriptive only.
MASAI trial numbers at a glance
Reported results from the three main MASAI publications:
- Lancet Oncology 2023 (80,033 women): 6.1 vs 5.1 cancers per 1,000; recall 2.2% vs 2.0%; false positives 1.5% in both groups
- Lancet Digital Health 2025 (105,934 women): 338 vs 262 cancers; 6.4 vs 5.0 per 1,000
- Invasive cancers 270 vs 217; in situ cancers 68 vs 45
- Workload: 61,248 vs 109,692 screen readings, 44.2% fewer
- The Lancet 2026: interval cancers 1.55 vs 1.76 per 1,000; rate ratio 0.88 (95% CI 0.65 to 1.18)
- Sensitivity 80.5% vs 73.8%; specificity 98.5% in both groups
How much did AI cut radiologists' workload?
Because most mammograms were sent to one radiologist instead of two, the AI group needed far fewer screen readings. In the 2023 safety analysis, radiologists did 46,345 readings in the AI group and 83,231 in the standard group, a 44.3% reduction. In the full trial, the figures were 61,248 and 109,692 readings, a 44.2% reduction.
This matters in many countries. Kristina Lång, who led the study, has described breast radiology as a speciality that is in short supply. One of the authors has suggested the approach could ease radiologist workload and might shorten waiting times for women.
Workload was measured as the number of readings. The studies did not measure time spent per image, extra checks of AI-flagged exams, or the cost of the software, so a 44% cut in readings is not the same as a 44% cut in cost or time.
False positives and recalls with AI mammography screening
A false positive is when a woman is called back for more tests and no cancer is found. Recalls cause worry, extra imaging and sometimes biopsies, so a screening method that finds more cancers by recalling many more women is not necessarily better.
In MASAI, this did not happen. In the full trial, 1,110 women were recalled in the AI group and 1,027 in the standard group, a difference that was not statistically significant. The false-positive rate was almost identical between groups. The share of recalled women who turned out to have cancer, called the positive predictive value, was higher with AI. In the 2023 analysis it was 28.3% with AI and 24.8% without.
Some questions remain. AI support found more in situ cancers, and the trial cannot yet say how many of these would have caused harm. Lång has said further studies on future screening rounds and on cost-effectiveness are needed to understand the long-term benefits and risks.
What do other studies of AI mammography screening show?
The PRAIM study in Germany, published in Nature Medicine in 2025, looked at 463,094 women screened by 119 radiologists at 12 sites. Radiologists chose whether to use an AI system for each mammogram, so this was an observational study, not a randomised trial. The detection rate was 6.7 per 1,000 with AI support and 5.7 per 1,000 without, 17.6% higher, while recall rates were similar (37.4 vs 38.3 per 1,000). The study was funded by Vara, the company that makes the AI, and several authors work for or have ties to the company.
In England, the government announced the EDITH trial on 4 February 2025. It plans to include nearly 700,000 women at 30 NHS sites, with £11 million of funding from the National Institute for Health and Care Research. It will test whether AI can help radiologists find early changes, and whether one specialist can safely do a reading that currently needs two.
These studies point in the same direction as MASAI, but they used different AI products, methods and health systems.
Limits of the AI mammography screening evidence
What the current evidence cannot yet tell us:
- MASAI was one trial in one region of Sweden, using one AI product (Transpara 1.7.0) and experienced breast radiologists.
- Radiologists knew when AI was in use, which could have changed how carefully they read.
- The drop in interval cancers (82 vs 93) was not statistically significant; it showed AI was not worse.
- No study yet shows that AI-supported screening lowers deaths from breast cancer.
- Results from later screening rounds and cost-effectiveness are still awaited.
- The German PRAIM study was observational and company-funded, so it is weaker evidence than a randomised trial.
What does AI mammography screening change for the woman being screened?
In the MASAI design, very little changes on the day. The woman has the same mammogram, with the same machine and the same positioning. The AI works in the background when the images are read, and a radiologist still looks at every exam and makes the decision.
The possible differences come later. If AI support finds a small cancer that would otherwise have been missed, she may be called back sooner. If the AI sorts her mammogram as low risk, it may be read by one radiologist instead of two. If her result is normal, the trial suggests her chance of an interval cancer is not higher with AI support.
The authors stress that the study does not support replacing health professionals with AI. Any recall, biopsy or treatment decision is still made by doctors, using the full set of images, examination findings and, where needed, pathology.
AI mammography screening evidence by setting
Where the evidence comes from, in brief:
- Sweden: MASAI randomised trial, about 106,000 women in an organised programme with double reading
- Germany: PRAIM observational study, 463,094 women, AI used at the radiologist's choice
- England: EDITH trial announced in February 2025, planned to include nearly 700,000 women
- India: public screening relies mainly on clinical breast examination, not routine mammography for all women
Is AI mammography screening relevant in India?
The MASAI results come from a setting quite different from most of India. Sweden regularly invites women aged 40 to 74 for mammography, and two radiologists read every image. A 2025 scoping review in BMC Cancer reported that clinical breast examination, where a trained health worker examines the breasts by hand, remains the mainstay of breast screening efforts in India, and that participation is low.
India also has its own trial evidence on that approach. A cluster randomised trial in Mumbai of 151,538 women aged 35 to 64, published in the BMJ in 2021, found that clinical breast examination every two years by trained health workers lowered the share of cancers found at stage III or IV from 47% to 37%.
So AI mammography screening is most relevant for women who are already having mammograms, for example at a hospital or diagnostic centre. Whether a centre uses AI, which product, and how its radiologists use it will vary, and the MASAI figures may not apply. Your doctor can explain which screening is suitable for you.
What happens next for AI mammography screening research?
The MASAI team has said that further studies on later screening rounds in the same women, and on cost-effectiveness, are needed to understand the long-term benefits and risks. Lund University reports that several Swedish regions have already started using AI-supported screening, and that AI support has begun in other parts of Europe.
In England, the EDITH trial, announced in February 2025, plans to test AI support in nearly 700,000 women at 30 NHS sites, including whether one specialist can safely replace two readers. Results from different countries, AI products and health systems will show whether the MASAI findings hold up more widely.
The big open question is whether finding more small cancers earlier leads to fewer deaths, and whether it also leads to treating some cancers that would never have caused harm. Answering that will take years of follow-up.
Questions to ask your doctor about AI mammography screening
- Is a screening mammogram suitable for me, at my age and with my history?
- Does this centre use AI to help read mammograms, and which software is it?
- Will a radiologist read my mammogram, and will one or two radiologists read it?
- If AI flags something, what further tests could follow?
- If my breasts are dense, does that change how my mammogram is read or what else is advised?
- How often should I be screened, and what symptoms should make me see a doctor between screenings?
Sources
1. Gommers J, 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, 2026. https://doi.org/10.1016/S0140-6736(25)02464-X 2. Hernström V, et al. Screening performance and characteristics of breast cancer detected in the Mammography Screening with Artificial Intelligence trial (MASAI). The Lancet Digital Health, 2025. https://doi.org/10.1016/S2589-7500(24)00267-X 3. Lång K, et al. Artificial intelligence-supported screen reading versus standard double reading in the Mammography Screening with Artificial Intelligence trial (MASAI): a clinical safety analysis. The Lancet Oncology, 2023. https://doi.org/10.1016/S1470-2045(23)00298-X 4. Lund University. AI support in breast cancer screening leads to fewer missed cancer cases. Lund University, 2026. https://www.lunduniversity.lu.se/article/ai-support-breast-cancer-screening-fewer-missed-cancer-cases 5. The ASCO Post. Randomized trial shows AI-supported mammography improves sensitivity and lowers interval cancer rate. The ASCO Post, 2026. https://ascopost.com/news/february-2026/randomized-trial-shows-ai-supported-mammography-improves-sensitivity-and-lowers-interval-cancer-rate 6. Eisemann N, et al. Nationwide real-world implementation of AI for cancer detection in population-based mammography screening. Nature Medicine, 2025. https://doi.org/10.1038/s41591-024-03408-6 7. UK Government, Department of Health and Social Care. World-leading AI trial to tackle breast cancer launched. GOV.UK, 2025. https://www.gov.uk/government/news/world-leading-ai-trial-to-tackle-breast-cancer-launched 8. Venghateri JB, et al. Low participation in cancer screening in India: a scoping review of breast and cervical cancer programs. BMC Cancer, 2025. https://pubmed.ncbi.nlm.nih.gov/41204154/ 9. Mittra I, et al. Effect of screening by clinical breast examination on breast cancer incidence and mortality after 20 years: prospective, cluster randomised controlled trial in Mumbai. BMJ, 2021. https://pubmed.ncbi.nlm.nih.gov/33627312/