Summary

• Vara’s Autonomous Triage has received Class IIb CE certification under the EU Medical Device Regulation for organised breast cancer screening.
• Mammograms classified as clearly normal can be reported without radiologist review, while other examinations remain in the human reading pathway.
• Adoption will still depend on national screening rules, monitoring, and evidence that autonomous performance remains dependable across different sites and populations.

Vara has received European regulatory certification for breast screening software that can report some mammograms as normal without a radiologist reading them, taking medical imaging AI beyond decision support and into a defined clinical task.

The Berlin company’s Autonomous Triage system has been certified as a Class IIb medical device under the EU Medical Device Regulation. When the software identifies an examination as clearly normal, that case can be reported without human review, while mammograms outside the autonomous pathway continue through conventional reading. Vara describes the certification as the first for breast screening AI operating in this way.

Most imaging AI currently assists clinicians by highlighting suspicious areas, prioritising examinations, or acting as a second reader. Autonomous triage changes the operating model because the software can complete part of the workflow rather than merely advise the professional responsible for it. The potential reduction in radiologist workload is therefore greater, but so is the need to monitor performance after deployment.

Certification does not mean the system can be switched on across every European screening service immediately. National screening guidelines, procurement, clinical practice, and local governance will determine how quickly programmes can adopt autonomous reporting, leaving a sizeable gap between regulatory clearance and routine use.

Monitoring becomes part of the system

Vara has paired the model with a monitoring system called ATMON, which tracks performance and operating conditions at individual sites. The company says it follows factors including mammography hardware, system health, cancer detection rates, recall rates, and daily performance signals, and can return a site to full radiologist reading when indicators move outside defined limits.

That architecture addresses a central problem in autonomous clinical AI. A model validated successfully in one environment can encounter different machines, populations, workflows, and disease prevalence when deployed elsewhere, while software updates and equipment changes can alter conditions over time. Once the technology is completing a clinical task rather than supporting one, detecting deterioration quickly becomes part of the safety case.

Vara already has large scale evidence for AI supported breast screening, although that evidence should not be confused with a trial of the newly certified autonomous pathway. The PRAIM implementation study, published in Nature Medicine in 2025, covered more than 460,000 women in Germany and compared AI supported double reading with conventional screening. Researchers reported a breast cancer detection rate of 6.7 per 1,000 in the AI supported group, compared with 5.7 per 1,000 without AI support, while recall rates did not increase.

More recent prospective work has also examined whether low risk mammograms can be removed from human reading under controlled conditions, adding evidence that selective automation can reduce workload. Even so, clinical adoption rests on several measures moving together: cancer detection, unnecessary recalls, workload, performance across imaging systems, and the ability to identify when the technology should no longer be trusted to act autonomously.

Healthcare AI moves beyond decision support

Healthcare makes the distinction between assistance and automation unusually visible because responsibility cannot be treated as an abstract software issue. A recommendation leaves a clinician clearly responsible for accepting or rejecting it, while a system that completes a defined clinical task requires organisations to decide who intervenes when monitoring detects drift, when equipment changes, or when an examination reported as normal later proves otherwise.

European healthcare providers will also have to fit autonomous systems into the wider regulatory environment around high risk AI. Medical device certification covers a crucial part of the framework, while risk management, monitoring, data governance, documentation, and human oversight continue to shape how the technology can be operated in practice.

The appeal for screening services is straightforward. Population programmes generate large volumes of normal examinations, while radiology capacity remains constrained across much of Europe. Removing a carefully selected proportion of those cases from human reading could release specialist time for ambiguous and suspicious examinations without eliminating professional judgement from the wider service.

The economics will vary sharply between health systems. A programme able to remove a material share of normal cases from the human reading queue may create useful capacity, while another operating under rules that retain mandatory review could capture far less of the labour saving. Integration, monitoring, training, and governance also carry costs that sit outside the headline performance of the model.

Vara says more than half of Germany’s organised breast screening programme already uses its technology, processing more than 250,000 screenings a month. That installed base gives the company a substantial environment in which to test whether autonomous use can move from certification into regular service delivery.

The next evidence will come from those deployments rather than another benchmark. If European screening programmes begin allowing clearly normal mammograms to leave the human reading queue, the safety of the system will depend as much on recognising when automation should stop as on the classifier deciding when it can proceed.