AI improves mammography screening overall, including in dense breasts, but it cannot find a cancer the mammogram does not show. In the MASAI trial, AI-supported screening raised sensitivity from 73.8% to 80.5%, consistently across breast density. For dense breasts, AI’s biggest role may be deciding who needs MRI or contrast mammography.
Around the world, women are being told something new after a mammogram: that their breasts are dense. In the United States, the Food and Drug Administration’s updated mammography rules have required facilities since September 2024 to tell every patient whether her breasts are dense or not dense, and that dense tissue can make cancer harder to find. It is useful information. It is also, for many women, an unsettling letter with no clear next step.
Dense breast tissue matters in two ways. It is an independent risk factor for breast cancer, and it masks tumours on a mammogram, because both dense tissue and cancer appear white on the image. Cancers that the screening mammogram misses and that surface between screens, interval cancers, are more common in women with dense breasts.
That makes dense breasts an obvious target for artificial intelligence, and a common hope follows: AI may help most where mammography is weakest. The argument here is that the hope is half right. AI does improve reading in dense breasts. But the larger gain probably comes from using AI to decide which women need a different kind of test, because no reader, human or machine, can find a tumour that is hidden in the pixels.
Does AI detect more cancers in dense breasts?
The best evidence comes from large European screening studies, where two radiologists typically read each mammogram.
The MASAI trial in Sweden randomised 105,934 women to AI-supported screening or standard double reading. In its interval cancer analysis, published in The Lancet, sensitivity was 80.5% with AI support against 73.8% with standard double reading, while specificity was 98.5% in both groups. Interval cancers, the trial’s key safety measure, were 1.55 per 1,000 in the AI arm and 1.76 per 1,000 in the control arm, meeting the test for non-inferiority. The authors reported that the sensitivity gain was consistent across age and breast density. The AI approach also cut screen-reading workload substantially, with 44.3% fewer total reads.
In Germany, the PRAIM study, published in Nature Medicine in 2025, followed 463,094 women screened in routine practice, 260,739 of them read with AI support. Detection was 6.7 cancers per 1,000 with AI against 5.7 without, a 17.6% increase, with a slightly lower recall rate.
These are strong real-world results. They suggest AI helps radiologists find cancers they would otherwise miss across the screening population, and that dense breasts are not excluded from that benefit. That is good news, and it should not be minimised.
But look closely at what AI is doing. It reads the same mammogram. If a cancer produces some faint signal on the image, AI may catch what a tired human overlooks. If the tumour is completely hidden behind dense tissue, no amount of better reading will reveal it.
Why can’t AI solve the masking problem on its own?
The answer lies in the evidence on supplemental imaging, which shows how much a mammogram misses in the densest breasts.
The DENSE trial in the Netherlands, published in the New England Journal of Medicine in 2019, enrolled 40,373 women aged 50 to 75 with extremely dense breasts and a normal mammogram. Women invited to supplemental MRI had an interval cancer rate of 2.5 per 1,000 screenings, half the 5.0 per 1,000 in the mammography-only group. Among women who actually had MRI, the cancer detection rate was 16.5 per 1,000 screenings, cancers that the mammogram had already missed. The cost was a false-positive rate of 79.8 per 1,000 MRIs.
The BRAID trial in the UK, published in The Lancet in May 2025, randomised women with dense breasts and a negative mammogram to one of three supplemental tests. Abbreviated MRI detected 17.4 cancers per 1,000 examinations and contrast-enhanced mammography 19.2 per 1,000, while automated breast ultrasound detected 4.2 per 1,000. MRI and contrast mammography found three times as many invasive cancers as ultrasound, and they were half the size.
Compare those numbers with the AI gains. PRAIM added about one cancer per 1,000 women screened across the whole population. Supplemental MRI or contrast mammography in women with dense breasts found roughly 16 to 19 extra cancers per 1,000 examined. The difference in magnitude reflects a physical limit. MRI and contrast mammography use a different signal, blood flow to tumours, that dense tissue does not hide. AI reading the standard mammogram works within the limits of that image.
How could AI decide who needs supplemental screening?
This is where AI’s role in dense breasts may be most valuable.
Supplemental imaging is expensive, capacity-limited and brings false positives. Offering MRI to every woman with dense breasts is beyond most health systems; close to half of women screened in the United States have dense breasts by commonly cited estimates, and MRI capacity is finite. The question is which of those women benefit most.
Density alone is a blunt instrument. Many women with dense breasts will never develop cancer, and some with less dense breasts are at higher risk. AI can help in two ways. Risk models that read the mammogram image itself can estimate a woman’s future cancer risk more accurately than density categories and traditional questionnaires, according to several retrospective studies. And AI can estimate the likelihood that a given mammogram is masking a cancer, based on the image’s appearance.
Combined, those tools could direct supplemental MRI or contrast mammography to the women most likely to benefit, rather than rationing it by density category or by who can pay. That is a use of AI that attacks mammography’s real weakness in dense breasts rather than trying to see past it. It is also still largely unproven in prospective trials. The risk models have strong retrospective data; the evidence that AI-guided triage to supplemental imaging reduces interval cancers or deaths is still being built.
A worked example: 10,000 women with dense breasts
Hypothetical, with round numbers drawn loosely from the trials above for illustration. A screening programme has 10,000 women with dense breasts and normal mammograms each round, and MRI capacity for 2,000 of them.
Option one: choose the 2,000 at random. Using detection rates around those in DENSE and BRAID, roughly 17 per 1,000, MRI finds about 34 additional cancers, with around 160 false positives needing further work-up.
Option two: use an AI risk and masking model to choose the 2,000 at highest predicted risk. If the model concentrates, say, half of all hidden cancers into that top 20%, MRI finds far more than 34 cancers from the same scanner time. The actual gain depends entirely on the model’s real-world performance, which is exactly what prospective trials need to measure.
Option three: skip supplemental imaging and rely on AI-assisted mammography reading. Based on current evidence, that improves detection somewhat across the population, but leaves most of the masked cancers in this group undetected until they surface as interval cancers.
The point of the example is not the specific numbers. It is that AI’s contribution may be larger as a triage tool for scarce supplemental imaging than as a better reader of an image that has already hidden the cancer.
What is the strongest case for AI reading alone?
The counterargument is practical. Supplemental MRI and contrast mammography are unavailable or unaffordable for most women in the world. AI-assisted mammography reading can be deployed wherever mammography exists, and the MASAI and PRAIM results show that it improves detection, reduces workload and does not increase interval cancers, including among women with dense breasts. In a system where the choice is AI-assisted mammography or standard mammography, the choice is easy.
That is correct, and it is an important caveat for health systems in lower-income settings, where mammography capacity itself is the constraint. AI reading also frees radiologist time, which can be redirected to supplemental imaging programmes. But the counterargument does not answer the core question for dense breasts: better reading of a masked image has a ceiling, and it is well below what a different imaging modality can achieve. Women told they have dense breasts deserve to know that. The follow-up problem matters too; as we argued in our piece on cancer screening gaps, a screening programme is only as good as its path from abnormal result to diagnosis.
What this means for screening programmes, radiologists and women
For screening programmes and health systems, AI-supported mammography reading now has strong trial and real-world evidence and can be adopted with appropriate monitoring. For dense breasts, the strategic question is supplemental imaging capacity and how to allocate it. Pilot AI-based risk and masking triage with prospective evaluation of interval cancers.
For radiologists, AI is a second reader or triage tool that can reduce workload and improve sensitivity. Dense breast cases still require judgment about additional imaging, and radiologists should understand what their AI tool’s performance looks like in dense tissue specifically, since overall figures can hide subgroup differences. For a broader view on why diagnostic AI performs best on narrow tasks, see our analysis of AI diagnostics evidence.
For policymakers, density notification without access to supplemental imaging creates anxiety without a pathway. Coverage decisions for supplemental MRI or contrast mammography, and investment in capacity, matter more than notification alone.
For vendors, report performance by density category, validate risk and masking models prospectively, and avoid implying that AI reading solves the dense breast problem.
For women, if you are told you have dense breasts, ask your doctor what that means for your personal risk and whether additional screening is appropriate for you. Continue attending routine screening.
Frequently asked questions
Does AI improve mammography for women with dense breasts?
Yes, to a degree. The MASAI trial reported in The Lancet that AI-supported screening increased sensitivity from 73.8% to 80.5%, consistently across breast density, without increasing interval cancers. But AI reads the same mammogram image, so it cannot detect cancers that dense tissue completely hides.
What is the best supplemental screening for dense breasts?
In the UK BRAID trial, published in 2025, abbreviated MRI detected 17.4 and contrast-enhanced mammography 19.2 extra cancers per 1,000 examinations in women with dense breasts and normal mammograms, compared with 4.2 for automated ultrasound. The Dutch DENSE trial found supplemental MRI halved interval cancers. Availability and cost vary; discuss options with your doctor.
Can AI decide who needs an MRI for dense breasts?
Potentially. AI risk models that analyse mammogram images, and models that estimate whether an image may be masking a cancer, could prioritise women for scarce supplemental imaging. Retrospective results are promising, but prospective trials showing reduced interval cancers are still needed before this becomes standard.
Why are dense breasts a problem for screening?
Dense tissue and tumours both appear white on a mammogram, so tumours can be hidden, and dense breasts are also an independent risk factor for breast cancer. This means more cancers are missed on screening and appear between screens. In the United States, facilities must now tell women whether their breasts are dense.
Should women with dense breasts still have mammograms?
Yes. Mammography still detects many cancers in dense breasts and remains the foundation of screening. Women with dense breasts should discuss with their doctor whether supplemental imaging is appropriate based on their overall risk, and continue routine screening as recommended.
AI makes a good mammogram reader better. In dense breasts, the problem is often not the reader but the picture. The biggest gain will come from using AI to decide who needs a better picture.
For radiologists and screening leads: are you using AI to read dense breasts, to triage them, or both, and what have you seen so far?
