Vara, a Berlin-based medical software company trading as MX Healthcare GmbH, has announced that it received the world's first CE mark for autonomous breast screening AI. The certificate allows the company's system to close screening examinations it classifies as clearly normal without a radiologist examining them. The announcement arrived Wednesday through a statement issued by the company's chief executive Jonas Muff. Muff also published a longer post explaining the design rationale and regulatory significance.

Why autonomous triage matters

European organised breast screening programmes routinely require every mammogram to be read twice. This is intended to reduce errors and catch subtle cancers. Vara says about 97 percent of screening images turn out to be normal. In that environment, using two human reads on every image becomes difficult to sustain. The radiologist workforce is shrinking in many European countries, and screening age ranges are widening. Autonomous triage addresses that pressure by allowing artificial intelligence to clear only the images it considers obviously normal, directing human attention to the smaller set of examinations that need interpretation.

What the certification permits

The approval is a Class IIb CE certification under the European Medical Device Regulation. Vara describes the approved function as autonomous triage in organised breast cancer screening. In this mode, the AI examines each study and reports those it classifies as clearly normal. No radiologist reads those cases. Every other examination is still sent to at least one human reader. The certification is limited to organised population screening programmes. Symptomatic patients and diagnostic work-up are not included.

A different category from computer-aided detection

The clearance goes beyond traditional computer-aided detection tools. Those tools have been sold since the early 2000s and mark a region or feature on an image for a doctor to inspect. The deep learning generation that followed provides a judgement about how suspicious a finding is. In both cases, the AI runs alongside the radiologist and does not replace the human read. Vara's certificate allows the system to replace that human read for a defined subset of cases. Screening programmes therefore need to think about it as a new category rather than as familiar decision-support software.

The safety system wrapped around the AI

Vara attributes the certification less to the model itself than to the safety layer built around it. The company calls that layer ATMON, which stands for Autonomous-Triage Monitoring. According to Vara, ATMON sets and monitors each site's operating point. It tracks changes in mammography hardware, system health and daily performance signals. It also monitors cancer detection rates and recall rates. If those numbers move outside defined limits, ATMON is intended to revert the site to full radiologist reading.

Stefan Bunk, Vara's chief technology officer and co-founder, said the case rested on seven years of continuous monitoring of how the AI performs in the real world on every single case. Muff compares the arrangement to an autopilot that hands control back for takeoff and landing. The monitoring layer is now being opened to all Vara customers. That includes organisations in countries whose national guidelines do not permit autonomous triage, and customers with no intention of adopting it at present.

Published evidence and its limits

The company's announcements rely on the PRAIM study, published in Nature Medicine in 2025, as the main evidence for its regulatory clearance. PRAIM screened 463,094 women at 12 German sites between July 2021 and February 2023. In the AI-supported group, 119 radiologists read 260,739 cases. Cancer detection came out at 6.7 per 1,000 women in the AI-supported group, compared with 5.7 per 1,000 in the control group, a relative increase of 17.6 percent. Recall rates were slightly lower in the AI-supported group.

Yet PRAIM tested AI-supported double reading. It did not test autonomous triage. Participating radiologists also chose for themselves whether to use the AI, and the paper records this limitation. That makes the comparison something other than a randomised trial. Vara's announcement calls the study nationwide and gives a cohort size of 461,818 women, a figure that the paper does not carry. These details do not necessarily mean the certification is unsound. A notified body assessed the product, although Vara has not named that body. Screening programmes weighing adoption should hold on to the distinction between the study endpoint and the approved product use.

A 2018 precedent outside breast imaging

Vara is not the first autonomous diagnostic AI in medicine, but it is the first in breast imaging. In April 2018, IDx-DR received authorisation from the United States Food and Drug Administration through the De Novo pathway for autonomous detection of diabetic retinopathy in retinal images. That system was designed for primary care settings where no clinician interprets the result. Vara's milestone is narrower and specific to breast imaging, a field where such a clearance had gone unclaimed for a decade.

Other autonomous or AI-native health products have also made progress recently. Flok Health received approval for an AI physiotherapist that the NHS lets work without direct supervision. French startup Raidium has built a radiology viewer with AI at its core. The same week, Healthwatch found AI scribes getting diagnoses wrong in NHS records, a reminder that automation in healthcare needs continuous checks.

The EU AI Act timeline

Vara presents ATMON as the kind of human oversight that the European Union's AI Act expects from high-risk systems. Its two statements, however, describe the legal situation differently. The company's main release says the Act will require this kind of monitoring. Muff's post says the Act requires it now. The technical timeline supports the release's wording. The AI Act treats AI systems placed inside medical devices regulated by the Medical Device Regulation as automatically high-risk. But obligations for this Annex I category, including the human oversight duty in Article 14, were deferred until 2 August 2028. Regulators already hold enforcement powers over general-purpose models, but medical devices are on a slower compliance clock. Vara is therefore early rather than obligated under the

Source:TNW | Health-tech News