Breast cancer AI is no longer just a second pair of eyes on a mammogram. A post on NVIDIA’s blog on 5 October, timed for Breast Cancer Awareness Month, profiles four startups in its Inception programme that apply AI at different points of the care pathway: getting a scan closer to the patient, helping radiologists read the growing pile of mammograms, predicting which treatments will work and planning surgery in 3D.
NVIDIA frames the problem bluntly. Breast cancer is the most commonly diagnosed cancer among American women, “yet the gaps in care are wide”. Many women skip screening, radiologists are reading more images with fewer colleagues, and the tests that guide treatment “can take weeks to return results”.
This article sets out those gaps with the latest figures, what each of the four companies does and where it stands with regulators, what large trials say about AI-supported screening, the limits worth keeping in mind, and what all of it means for the NHS and for UK health-tech builders.
Table of contents
- The Gaps Breast Cancer AI Is Trying to Close
- Breast Cancer AI at the Scan: iSono Health
- Breast Cancer AI in the Reading Room: Whiterabbit.ai
- What Trials Say About Breast Cancer AI in Screening
- Breast Cancer AI for Treatment Decisions: Ataraxis AI
- Breast Cancer AI in Surgery Planning: SimBioSys
- The Four Breast Cancer AI Companies Side by Side
- Limits and Open Questions for Breast Cancer AI
- Questions to Ask Before Adopting Breast Cancer AI
- What Breast Cancer AI Means for the UK and NHS
- Breast Cancer AI FAQ
- References
The Gaps Breast Cancer AI Is Trying to Close
Every breast cancer AI product targets a specific bottleneck. It helps to size those bottlenecks before looking at the tools.
The scale of the problem
The American Cancer Society’s 2026 estimates put the burden in plain numbers.
| US measure (ACS, 2026) | Estimate |
|---|---|
| New invasive breast cancers in women | About 321,910 |
| New ductal carcinoma in situ (DCIS) cases | About 60,730 |
| Deaths from breast cancer in women | About 42,140 |
| Lifetime risk for a US woman | About 13%, or 1 in 8 |
| Fall in death rate, 1989 to 2023 | 44% |
| Median age at diagnosis | 63 |
The 44% fall in deaths since 1989 is largely credited to earlier detection and better treatment. The gaps NVIDIA highlights are the places where that progress stalls.
Missed screening
NVIDIA links to a MedStar Health survey which found that 60% of women aged 40 and over were not having the yearly mammogram the American College of Radiology recommends for women at average risk, and 21% had never had one. The most common reason given was a history of normal results, followed by pain, cost and lack of insurance.
One nuance: US guidance is not unanimous on frequency. The US Preventive Services Task Force’s final 2024 recommendation is screening every other year from 40 to 74, while the American College of Radiology recommends annual screening. “Missing the annual mammogram” therefore covers some women who are following the federal schedule.
Fewer radiologists, more mammograms
NVIDIA cites about 40 million US mammograms a year. The FDA’s own register is higher: as of 1 September 2026, facilities reported 45,025,781 annual mammography procedures across 9,128 certified facilities, of which 8,657 have tomosynthesis (3D mammography) units. NVIDIA also points to a projected shortfall of tens of thousands of radiologists over the next decade. In the UK, the Royal College of Radiologists’ 2025 census found a 32% shortfall of consultant radiologists, more than 2,300 full-time equivalents, forecast to reach 40% by 2030.
Waiting weeks for treatment tests
At the other end of the pathway, decisions about chemotherapy often depend on genomic assays sent to outside labs. NVIDIA says these can involve “a two- to four-week wait” for a separate tissue test, at the moment patients most want certainty. That delay is the gap that treatment-side breast cancer AI is built to close.
Breast Cancer AI at the Scan: iSono Health
The first company in NVIDIA’s post works on access, the gap before a radiologist ever sees an image.
Two minutes instead of 45
iSono Health’s ATUSA is a wearable, automated 3D ultrasound system. NVIDIA says it captures a standardised breast volume in about two minutes per breast, against up to 45 minutes for a conventional handheld ultrasound.
Two minutes is about 4.4% of 45 (2 divided by 45), a time saving of up to roughly 95% per breast. The comparison is against the upper end of handheld scan times, so the typical saving will be smaller.
Repeatable scans over time
Handheld ultrasound depends on whoever holds the probe, so one year’s scan rarely lines up with the next. ATUSA captures the whole breast the same way every time. NVIDIA says the system’s AI was trained on thousands of full-breast scans comprising more than 1.5 million ultrasound frames, and that iSono describes the 3D scan as 28% more sensitive than handheld 2D ultrasound. That figure is the company’s claim, not a trial result.
Clearance, availability and the next study
iSono announced FDA clearance for ATUSA in May 2022. NVIDIA says it is commercially available through partner clinics in California, Texas, Georgia, Tennessee and Washington DC. A multicentre study of 3,200 patients is under way, led from UC Davis and Vanderbilt University Medical Center, and the company plans to extend its breast cancer AI across 3D ultrasound, mammography, MRI and clinical data.
Breast Cancer AI in the Reading Room: Whiterabbit.ai
The second company targets the radiologist workload itself.
Density and long-term risk
Whiterabbit.ai’s FDA-cleared WRDensity software assesses breast density from mammograms automatically and, NVIDIA says, has been used in the care of hundreds of thousands of patients. Dense tissue both raises cancer risk and makes tumours harder to see on a mammogram. The company also offers WRRisk, decision-support software that estimates a patient’s long-term risk of developing breast cancer.
One cancer in every 200 mammograms
“Every day, breast radiologists face a needle-in-a-haystack problem, trying to find roughly one cancer in every 200 mammograms,” said Jason Su, Whiterabbit’s cofounder and chief technology officer. One in 200 is 0.5%, which means 199 of every 200 reads are, in the end, cancer-free. That ratio is why breast cancer AI that can confidently set aside clearly normal scans could free a large share of reading time.
Automating negative reads is still research
Whiterabbit is researching a new generation of AI that could help radiologists detect more cancers while automating the screening of mammograms that are negative. NVIDIA’s post ends with a caveat that “certain technologies described in this article are investigational and have not been approved by the U.S. FDA for commercial use”. Autonomous clearing of normal scans falls in that category today. Whiterabbit trains on NVIDIA GPUs at Washington University in St. Louis and runs inference on GPUs in the clinic.
What Trials Say About Breast Cancer AI in Screening
Vendor claims are a starting point. The strongest evidence for breast cancer AI in screening comes from two large European studies.
MASAI in Sweden
The MASAI trial, published in The Lancet Digital Health in 2025, randomised 105,934 women in Sweden’s national programme to AI-supported screening or standard double reading. AI-supported screening detected 6.4 cancers per 1,000 women against 5.0 in the control group, a 29% increase, without more false positives. Radiologists’ screen-reading workload fell by 44%.
PRAIM in Germany
The PRAIM study, published in Nature Medicine in January 2025, followed 463,094 women screened at 12 German sites by 119 radiologists. Radiologists using AI support detected 6.7 cancers per 1,000 against 5.7 without it, 17.6% higher, while the recall rate was slightly lower at 37.4 per 1,000 against 38.3.
Bar widths are each rate divided by 6.7. The relative gains check out against the stated rates: 6.4 divided by 5.0 is 1.28, and 6.7 divided by 5.7 is 1.175, matching the published 29% and 17.6% after rounding.
EDITH in England
England is now testing the same idea at scale. The EDITH trial, announced in February 2025 with £11 million of National Institute for Health and Care Research funding, is expected to involve almost 700,000 women at around 30 sites. It tests whether AI can take the place of one of the two human readers the NHS uses for each screen. Results will shape whether breast cancer AI becomes routine in NHS screening.
What the trials do not settle
Both European studies used AI to support, not replace, radiologists, and both measured detection rather than deaths. Detecting more small cancers is good only if it improves outcomes without adding overtreatment, and that takes years of follow-up. US screening, with one reader and tomosynthesis, also differs from European double reading, so results may not transfer directly.
Breast Cancer AI for Treatment Decisions: Ataraxis AI
The third company works after diagnosis, where the question becomes which treatment to give.
Predictions from slides already on file
Ataraxis AI builds models that predict patient outcomes and response to therapy from digital data, including pathology slides that are already part of a standard workup. The aim is to avoid a separate tissue test and its wait. “The tools oncologists rely on today to guide therapy decisions were largely trained once, fifteen years ago, and never updated,” said Joseph Cappadona of Ataraxis.
Two models, two decisions
NVIDIA describes two models. One predicts whether chemotherapy before surgery is likely to shrink a tumour enough to count as a response. The other, after surgery, estimates five-year recurrence risk and the likely benefit of chemotherapy. NVIDIA says both have been validated across more than 10 institutions and multiple clinical trials and are in active clinical use, running on NVIDIA GPUs with PyTorch and CUDA.
Why this part of breast cancer AI matters
Treatment-side breast cancer AI could matter as much as screening tools. Avoiding chemotherapy that would not help spares patients serious side effects, and getting an answer from existing slides removes weeks from the pathway. The evidence bar is high, though: oncologists will want prospective data showing these predictions change decisions for the better, not just that they correlate with outcomes.
Breast Cancer AI in Surgery Planning: SimBioSys
The fourth company turns scans into 3D models for surgeons.
3D tumour models from MRI
SimBioSys builds 3D models of breast tumours, veins and soft tissue from standard breast MRI to help guide surgery and treatment planning. Its TumorSight Viz platform received FDA 510(k) clearance in December 2023, and AuntMinnie reported a third clearance for version 1.3, which adds faster AI segmentation and PACS connectivity.
Multimodal recurrence risk
The company has also built a tool to estimate recurrence risk from 3D volumetric MRI data, pathology and clinical information. “We’re building a platform that now allows us to take multimodal data — imaging exams, pathology results, genomic testing when applicable and other biological inputs — and, using AI, bring that all together,” said chief executive Stacey Stevens at an NVIDIA panel in Phoenix on 1 October. SimBioSys uses NVIDIA MONAI and CUDA-X libraries on cloud GPUs.
The Four Breast Cancer AI Companies Side by Side
The four startups cover the pathway from first image to operating theatre. Their regulatory status varies by product, which matters more than the GPU stack behind them.
| Company | Stage of care | Product | US regulatory status | NVIDIA stack |
|---|---|---|---|---|
| iSono Health | Imaging access | ATUSA wearable 3D ultrasound | FDA cleared (2022); AI features expanding | GPU acceleration, open-source imaging tools |
| Whiterabbit.ai | Screening reads | WRDensity, WRRisk | WRDensity FDA cleared; negative-read automation in research | GPU cluster for training, in-clinic GPU inference |
| Ataraxis AI | Treatment choice | Response and recurrence models | In clinical use, per NVIDIA | PyTorch on CUDA, on premises and cloud |
| SimBioSys | Surgery planning | TumorSight Viz | FDA 510(k) cleared (2023 onwards) | MONAI, MONAI Deploy, cuBLAS, cloud GPUs |
The table reflects NVIDIA’s post plus each clearance announcement. Treat any feature not named in a clearance as investigational, as NVIDIA’s own disclaimer advises.
Limits and Open Questions for Breast Cancer AI
The promise is real, and so are the caveats. Four stand out.
A vendor post is not a trial
NVIDIA’s article is a showcase for companies in its startup programme, and several of its figures, such as the 28% sensitivity gain, come from the companies themselves. Independent trials like MASAI and PRAIM are the standard any breast cancer AI claim should eventually meet.
Cleared does not mean everything is cleared
A device can hold FDA clearance for one function while newer AI features remain investigational. Buyers should ask exactly which indication a clearance covers, on which equipment and for which population, before relying on a feature in practice.
Dense breasts and fairness
The USPSTF found the evidence insufficient to judge supplemental ultrasound or MRI screening for women with dense breasts. Tools such as automated ultrasound and density scoring sit right in that gap. Models also need testing across ages, ethnicities and breast densities, because a breast cancer AI trained on one population can underperform on another.
Who is accountable
When AI clears a scan that later proves to show cancer, responsibility has to sit somewhere. Clear protocols for how radiologists use, override and audit AI outputs matter as much as accuracy. That is why the European trials kept a human reader in every pathway.
Questions to Ask Before Adopting Breast Cancer AI
For a hospital, screening service or private clinic, the useful question is not whether breast cancer AI works in general but whether a specific product will work in your setting. Six questions cover most of the risk.
Which indication is cleared?
Ask for the exact regulatory clearance and its intended use: which modality, which equipment, which patient group and whether the output is advisory or triage. In the UK, also ask about UKCA or CE marking and any NHS assessment. A feature outside the cleared indication is research, however polished the demo.
Who was it tested on?
Ask for performance broken down by age, breast density, ethnicity and scanner vendor. A breast cancer AI model validated mainly on one population or one manufacturer’s machines can lose accuracy elsewhere. The MASAI and PRAIM results came from organised European programmes, which is a strength and also a limit.
How does it fit the reading workflow?
Decide in advance whether the AI acts as a second reader, a triage tool or a safety net, and what radiologists do when they disagree with it. The gains in the European trials came from carefully designed workflows, not from the software alone. Training time and the effect on reporting speed belong in the business case.
How will performance be monitored?
Breast cancer AI can drift when scanners, software versions or the screened population change. Agree how detection rates, recall rates and interval cancers will be tracked after go-live, who reviews them and what triggers a pause. A supplier that cannot support that audit trail is a risk in its own right.
Where does the data go?
Find out whether images leave your network, where inference runs, how long data is kept and whether it is used to retrain the model. Several of the companies NVIDIA profiles run inference on GPUs inside the clinic, which can simplify data protection.
What does it cost per study?
Pricing models range from per-study fees to site licences and revenue shares. Compare the total cost with the radiologist time it could release and the downstream costs of recalls. The value of breast cancer AI depends as much on that arithmetic as on its accuracy.
What Breast Cancer AI Means for the UK and NHS
The UK has a national screening programme, a radiologist shortage and an active AI trial, which makes it one of the places breast cancer AI will be tested hardest.
Screening volumes and workforce
In 2024/25 the NHS breast screening programme in England screened about 2.15 million women aged 45 and over, while the radiologist workforce shortfall stands at 32%. The NHS double-reads every screen, so if EDITH shows AI can safely replace one reader, it would free a large share of specialist time for diagnostic work and shorter waits.
For health-tech builders
For companies building in this space, the NVIDIA examples show the practical pattern: train on large imaging sets, run inference close to the clinic, and design for regulators from day one. Our guide to medical imaging software development covers that path, and our piece on private AI for healthcare explains why many clinics keep patient data on their own hardware.
Data, privacy and procurement
Health data is among the most sensitive categories under UK GDPR, and AI suppliers to the NHS face procurement and assurance checks before deployment. Clear data governance, audit trails and cybersecurity controls are prerequisites, not extras. For broader context on where clinicians see AI fitting, see our report on doctors weighing what AI leaves for them, and our data science services for building models responsibly.
Breast Cancer AI FAQ
What is breast cancer AI?
Breast cancer AI refers to software that uses machine learning to help with screening, diagnosis, risk assessment, treatment planning or surgery for breast cancer, usually by analysing images such as mammograms, ultrasound, MRI or pathology slides.
Does AI find more breast cancers?
In large European studies, yes. MASAI found 29% more cancers with AI support and PRAIM 17.6% more, without increasing false positives or recalls. Long-term effects on deaths are not yet known.
Will AI replace radiologists?
Not on current evidence. Trials use AI to support radiologists or replace one of two readers, with humans still making final decisions. England’s EDITH trial is testing the single-reader model.
Which companies did NVIDIA highlight?
iSono Health (wearable 3D ultrasound), Whiterabbit.ai (density, risk and reading support), Ataraxis AI (treatment response and recurrence prediction) and SimBioSys (3D surgical planning), all members of NVIDIA’s Inception startup programme.
Is breast cancer AI used in the NHS?
Some NHS trusts already use AI tools in breast imaging, and the national EDITH trial involving almost 700,000 women is testing AI in routine screening. Wider rollout depends on its results.
How often should women be screened?
In the US, the USPSTF recommends a mammogram every two years from 40 to 74, while the American College of Radiology recommends yearly screening. In England, the NHS invites women aged 50 to 71 every three years. Women should follow the advice of their own clinician.
References
NVIDIA Blog: From Scan to Treatment Plan, AI Helps Close Breast Cancer’s Deadliest Gaps
American Cancer Society: Key statistics for breast cancer
MedStar Health: Majority of women over 40 missing annual mammograms
USPSTF: Breast cancer screening final recommendation
Royal College of Radiologists: 2025 clinical radiology workforce census
The Lancet Digital Health: MASAI trial screening performance
Nature Medicine: Nationwide real-world implementation of AI in mammography screening (PRAIM)
Business Wire: FDA clears iSono Health’s ATUSA
AuntMinnie: FDA clears latest version of SimBioSys’s breast cancer platform
AuntMinnie: Whiterabbit gets FDA nod for breast density AI software
GOV.UK: Breast Screening Programme, England, 2024-25
PublicTechnology: NHS to trial AI in 700,000 breast cancer scans
