Contents
Of everything promised about artificial intelligence in medicine, one case has completed the full course: randomised trial, prespecified primary endpoint, published result. It is breast cancer screening, and the trial is called MASAI. With 105,934 women randomised within Sweden’s national programme, AI-supported reading detected 29% more cancers than standard double reading, without increasing false positives, and cut reading workload by 44%. And in 2026 came the missing figure: interval cancers, the endpoint that actually measures whether a screening programme is safe. This article explains why this case counts and what sets it apart.
The 30-second summary: the seven figures that matter
Each row names the study and the year. Note that the first three are successive publications from the same trial.
| Finding | Result | Study and year |
|---|---|---|
| Cancers detected per 1,000 women screened | 6.4 versus 5.0; relative proportion of 1.29 (95% CI: 1.09-1.51; p = 0.0021) | Hernström et al., Lancet Digital Health (2025) |
| False positives | No elevation: ratio standing at 1.01 (95% CI: 0.91-1.11; p = 0.92) | Hernström et al., Lancet Digital Health (2025) |
| Radiologist reading workload | 44.2% reduction in scans examined | Hernström et al., Lancet Digital Health (2025) |
| Interval cancers, the primary endpoint | 1.55 compared to 1.76 per 1,000; proportion of 0.88, remaining non-inferior | Gommers et al., The Lancet (2026) |
| Screening sensitivity | 80.5% versus 73.8% (p = 0.031), maintaining identical specificity | Gommers et al., The Lancet (2026) |
| Real-world rollout in Germany, 463,094 women | 6.7 against 5.7 per 1,000: marking a 17.6% increase in detection | Eisemann et al., Nature Medicine (2025) |
| Reading reduction across 31 reviewed studies | Ranging from 40% to 90% via triage methods | Jassim et al., BMJ Open (2025) |
What you will discover here
• How MASAI is designed and why its design differs from the rest.
• What an interval cancer is and why it was the missing figure.
- What transpired when it was administered to 463,094 women beyond the scope of a clinical trial.
• What the systematic review says is still unresolved.
The trial: how it was done
MASAI was conducted as part of the national screening initiative in Sweden, taking place across four distinct facilities situated in the country’s south-western region. During the timeframe spanning from April 2021 through December 2022, 105,934 female participants were allocated at random in a 1:1 ratio to undergo either artificial intelligence-assisted evaluation or conventional double reading devoid of AI. The median participant age stood at 53.7 years. Assignment was carried out automatically via the image archiving platform immediately following acquisition, ensuring that the interpreting radiologists remained blinded to the group assignment of each individual case.
The tool is used in two ways, and it is worth distinguishing them because that is where the result comes from. The system assigns each examination a risk score on a ten-level scale, used to decide whether that examination goes through one reading or two; and it also marks suspicious findings for the reading radiologist. At no point does it replace anyone: it reorganises the work and it points.
The detection results, published in Lancet Digital Health in 2025: 6.4 cancers per 1,000 women screened in the AI group against 5.0 in the control group, a ratio of 1.29 (95% CI: 1.09-1.51; p = 0.0021).
What kind of cancers were additionally detected
This question is what separates a useful screening programme from one that merely generates diagnoses. The extra detection largely involved small invasive tumours lacking nodal involvement: 58 additional T1 cancers, 46 more without lymph-node spread, and 21 more belonging to non-luminal A subtypes. In situ carcinomas also rose, with roughly half of that surge being high grade.
And what did not happen: neither the recall rate nor the false-positive rate rose significantly, while the positive predictive value of recall improved with a ratio of 1.19 (95% CI: 1.04-1.37; p = 0.012). Reading workload fell from 109,692 to 61,248 reads: 44.2% fewer.
| Study | What it contributed | Result | What it cannot support |
|---|---|---|---|
| Lång et al., Lancet Oncology (2023) | Safety analysis with the first 80,033 women | Similar or higher detection; 44.3% fewer reads | An interim safety analysis, not the final endpoint |
| Hernström et al., Lancet Digital Health (2025) | Secondary outcomes with the full sample | 29% more detection without more false positives | Detection is not the same as patient benefit |
| Gommers et al., The Lancet (2026) | Primary endpoint: interval cancer | Non-inferior, with higher sensitivity and equal specificity | It does not measure mortality; follow-up is two years |
| Eisemann et al., Nature Medicine (2025) | Real-world rollout at 12 German centres | 17.6% more detection without worsening recall | Radiologists voluntarily chose whether to use the tool |
| Jassim et al., BMJ Open (2025) | Review of 31 studies and over 2 million examinations | Reading reduction of 40% to 90% with triage | Interval-cancer follow-up incomplete in several datasets |
The missing figure: interval cancers
Identifying a higher volume of cancers doesn’t automatically prove the superiority of a screening initiative. An initiative can boost its detection stats by spotting growths that would have otherwise remained harmless, all while failing to catch the fast-moving ones. The metric designed to capture this flaw is the interval cancer: precisely the one surfacing in the gap between successive screening sessions, meaning the one the process failed to spot.
That constituted the main goal of MASAI, featured in The Lancet in 2026, utilizing a predefined non-inferiority margin of 20%.
| Metric | With AI | Standard double reading |
|---|---|---|
| Interval cancers per 1,000 participants | 1.55 (95% CI: 1.23-1.92) | 1.76 (95% CI: 1.42-2.15) |
| Proportion ratio | 0.88 (95% CI: 0.65-1.18) | p = 0.41: non-inferior |
| Invasive interval cancers | 75 | 89 |
| Interval cancers at stage T2 or above | 38 | 48 |
| Sensitivity | 80.5% (76.4-84.2) | 73.8% (68.9-78.3); p = 0.031 |
| Specificity | 98.5% | 98.5%; p = 0.88 |
The result meets the requirement and adds something more. Not only were there no additional interval cancers, but those that did occur in the AI group tended to be less invasive and at a lower stage, and programme sensitivity rose by nearly seven points without losing a tenth of a point of specificity. It is a clean result, and a rare one in this field.
What occurs beyond the trial
A clinical trial provides a controlled setting. Subsequently, whether this holds true within an actual programme remains to be seen, and a substantial reply is available. The PRAIM study, published in Nature Medicine in 2025, followed 463,094 women aged 50 to 69 screened at 12 German centres between July 2021 and February 2023, read by 119 radiologists.
The detection rate in the AI-supported group was 6.7 per 1,000 against 5.7 in the control group: 17.6% more (95% CI: 5.7% to 30.8%), statistically superior. The recall rate was slightly lower and non-inferior, and the positive predictive value rose from 14.9% to 17.9%.
There is one limitation the authors themselves flag and that should not be glossed over: radiologists voluntarily chose whether to use the system, so this is not random assignment and there could be differences between those who used it and those who did not.
What the systematic review qualifies
The broadest synthesis available was published by Ghufran Jassim and colleagues in BMJ Open in 2025: 31 studies covering more than two million screening examinations across Europe, Asia, North America and Australia.
Their conclusions match the above on the essentials. As a second reader or within double-reading workflows, AI maintains or modestly increases sensitivity — by up to nine percentage points — while preserving or improving specificity; and triage configurations reduce reading volume by between 40% and 90% while maintaining non-inferior detection when thresholds are conservatively calibrated.
But they add two reservations. The first concerns overall certainty: heterogeneity across designs and the use of enriched datasets limit some accuracy estimates, and interval-cancer follow-up is incomplete in several major studies. The second concerns autonomous use: transparent reporting, standardised evaluation and long-term population studies are needed before considering AI as a stand-alone reader.
Practical applications for this
Ordered by the strength of the evidence behind it:
• Always ask for the interval-cancer figure. It is the endpoint that separates detecting more from detecting better, and the one MASAI took three years to publish.
• The part illustrated here involves triage and support rather than acting as the sole reader. That approach is what the clinical trial assessed and what the systematic review currently deems safe.
• Look at what kind of tumour is additionally detected. In MASAI it was mostly small invasive cancers without nodal involvement, which is exactly what screening is for.
• Check that recall does not rise. The value of the result lies in detection increasing at no cost in false positives.
• Distinguish trial from rollout. The German study of 463,094 women is large, but use of the tool was voluntary and that is not random assignment.
Frequently asked questions
Does artificial intelligence detect more breast cancers than radiologists?
In supported screening, indeed, this is quantified via a randomised trial. Within MASAI, encompassing 105,934 women derived from Sweden’s national initiative, the detection rate reached 6.4 malignancies per 1,000 evaluations utilising AI, compared to 5.0 through conventional double reading, yielding a ratio of 1.29 (95% CI: 1.09-1.51; p = 0.0021). The artificial intelligence system did not operate autonomously; rather, it triaged examinations and highlighted suspicious areas for the specialist.
Do false positives increase?
Not significantly. Within that very same trial, the false-positive ratio stood at 1.01 (95% CI: 0.91-1.11; p = 0.92) alongside a recall ratio of 1.08 (0.99-1.17; p = 0.084), whereas the positive predictive value concerning recall saw an enhancement with a ratio reaching 1.19 (p = 0.012).
What is an interval cancer and why does it matter so much?
It is the cancer that appears between two screening rounds, that is, the one screening did not detect. It was MASAI’s primary endpoint, and the result published in The Lancet in 2026 was 1.55 per 1,000 with AI against 1.76 without it, a ratio of 0.88 (95% CI: 0.65-1.18) meeting the non-inferiority criterion. Without that figure, detecting more cancers does not show that screening is better.
How much work does it save?
Around half the reads. MASAI had 61,248 reads in the AI group against 109,692 in the control group, 44.2% fewer, and the review of 31 studies puts the reduction between 40% and 90% depending on the triage strategy used.
Is it possible for artificial intelligence to independently read mammograms?
Not yet, on the available evidence. The systematic review published in BMJ Open in 2025 concludes that transparent reporting, standardised evaluation and long-term population studies are needed before considering AI as a stand-alone reader, and notes that interval-cancer follow-up is incomplete in several of the available datasets.
What we do not know
• Whether it reduces mortality. No study has measured it: the hardest endpoint available so far is interval cancer at two years.
• What happens across several screening rounds. The published follow-up covers one interval, not a lifetime of successive screens.
- Whether the outcome relies on the specific setup. MASAI utilized a distinct release of a commercial software, and comparative, randomised evaluations between platforms are absent.
• How it behaves in different populations. The large studies are Swedish, Danish and German, with organised screening programmes and relatively homogeneous populations.
References
1. Hernström, V., Josefsson, 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. DOI: 10.1016/s2589-7500(24)00267-x
2. Gommers, J., Hernström, V. et al. Interval cancer, sensitivity, and specificity comparing AI-supported mammography screening with standard double reading without AI in the MASAI study. The Lancet, 2026. DOI: 10.1016/s0140-6736(25)02464-x
3. Lång, K., Josefsson, V. et al. Artificial intelligence-supported screen reading versus standard double reading in the MASAI trial: a clinical safety analysis. The Lancet Oncology, 2023. DOI: 10.1016/s1470-2045(23)00298-x
4. Eisemann, N., Bunk, S. et al. Nationwide real-world implementation of AI for cancer detection in population-based mammography screening. Nature Medicine, 2025. DOI: 10.1038/s41591-024-03408-6
5. Jassim, G., Otoom, O. et al. Performance of artificial intelligence in breast cancer screening programmes: a systematic review. BMJ Open, 2025. DOI: 10.1136/bmjopen-2025-111360
This content is for information only and does not constitute medical advice. The results described come from organised screening programmes using specific systems and protocols, and are not transferable to other care settings or to individual decisions about diagnostic tests.
