Vara Wins World-First CE Mark for Autonomous AI in Breast Cancer Screening: The Machine Now Reads the Normals

For a decade, everyone serious about AI in breast cancer screening knew where this was heading. Population screening programs read every mammogram twice, even though roughly 97% of those exams turn out to be normal. Under a worsening global radiologist shortage, double-reading every image is a formula for burnout, backlogs, and eventually unsustainable programs. Sooner or later, an AI system would have to report a share of the clearly normal mammograms on its own — the open question was whether it could do so safely enough to convince a regulator.

That question now has an answer. On September 2, 2026, Berlin-based Vara (MX Healthcare GmbH) announced it has received CE certification under the EU Medical Device Regulation as a Class IIb device for autonomous triage in organized breast cancer screening — the first authorization of its kind anywhere in the world. For examinations the system classifies as clearly normal, the AI can generate the report with no radiologist reading the case at all. Every other exam — anything ambiguous, suspicious, or technically difficult — still goes to a human reader, exactly as before.

What “autonomous” actually means here

The word “autonomous” gets used loosely in radiology AI, so it is worth being precise about what changed. Most AI tools in the field began as CADe (computer-aided detection): systems that mark a region for the radiologist to inspect, without judging how suspicious it is. The deep-learning generation moved the industry to CADx (computer-aided diagnosis), which marks findings and assigns them a level of suspicion. Nearly all modern breast AI, including Vara’s own flagship product, operates at this level — running alongside the radiologist as a concurrent second reader, not in place of one.

An autonomous AI is categorically different: it reports a case in place of a radiologist. There is a precedent in diagnostic imaging — IDx-DR, the diabetic-retinopathy system that became the first FDA-authorized autonomous AI in 2018 — but breast imaging had never crossed that line, despite being one of the most commercially developed corners of medical AI. Vara’s new certification opens the category. The company likens the arrangement to an autopilot: the system handles the long, stable stretches of reporting normals, while the safety layer tells radiologists when to take the controls back.

The evidence base: PRAIM

European regulators did not authorize this on benchmarks alone. The certification rests on the prospective PRAIM study (NCT04778670), published in Nature Medicine in January 2025 — the largest prospective AI study in healthcare to date. PRAIM followed 461,818 asymptomatic women aged 50–69 across 12 screening sites in Germany, comparing AI-supported screening against the standard double-reading protocol.

The headline result: radiologists working with AI achieved a breast cancer detection rate of 6.7 per 1,000 screens versus 5.7 per 1,000 in the control group — a relative increase of 17.6% (95% CI +5.7% to +30.6%) — without an increase in false-positive recalls. In the AI-supported arm, the detection rate was non-inferior and, by some analyses, statistically superior to unaided double reading. That combination — more cancers found, no more false alarms, and substantially less work per screen — is what made a serious conversation about autonomy possible.

Vara’s operational footprint made the real-world case even more concrete. More than 60% of Germany’s organized breast screening program now runs on Vara’s platform, processing over 250,000 screenings per month. The company has tracked every case put into clinical use since its first German deployments in 2019.

ATMON: the safety system that made it certifiable

According to Vara, the obstacle to certifying autonomous triage was never raw model performance. Two harder problems had to be solved first.

The first is model shift: even an excellent model’s performance can drift once it leaves the controlled conditions of a study — new mammography hardware, new sites with different workflows, a changing screening population as European programs extend screening ages. What matters is how the system behaves in the real world continuously, not how it scored once on a fixed dataset.

The second is the EU AI Act, which requires a human to stay on the loop for high-risk applications — and requires that oversight be engineered into the system, not bolted on afterward.

Vara’s answer to both is ATMON (Autonomous-Triage Monitoring), the real-time supervision layer the company says played a foundational role in the certification. ATMON sets and monitors each site’s operating point, tracks changes in mammography hardware, watches system health and daily performance signals, and follows cancer detection and recall rates. Whenever those signals move outside predefined limits, ATMON automatically reverts the site to full radiologist reading of all cases. It is built on seven years of continuous, per-case monitoring of Vara’s AI in production — precisely the kind of longitudinal evidence a notified body can audit. Notably, Vara says it will also license ATMON independently, so other providers deploying different autonomous screening AIs could integrate the same safety scaffolding.

Why this matters beyond mammography

The immediate effect is narrower than the headline suggests, and Vara is candid about that: screening programs will not switch on autonomous triage tomorrow. Rollout across Europe is subject to national screening guidelines and clinical readiness, and the certification’s larger significance is that it establishes a regulatory framework that previously did not exist. Until now, no AI system occupied the autonomous category in breast imaging, so screening guidelines had nothing to adopt. Now they do — which changes what prospective real-world research on autonomous systems can even look like.

The precedent also travels. The ATMON blueprint — continuous site-level monitoring, automatic reversion to human control on anomaly, hardware-drift tracking — is a generalizable answer to the EU AI Act’s human-oversight requirement for high-risk medical AI, and arguably a template for autonomous AI in any safety-critical domain. For an industry still arguing about whether “human in the loop” is a real safeguard or a fig leaf, a certified system that demonstrably hands control back when its own performance drifts is a concrete reference point.

The macro context matters too. Analysts at GlobalData expect AI-enabled digital health to help drive roughly 6% compound annual growth in Europe’s oncology digital-transformation segment through 2035, sustained by a steady drumbeat of regulatory approvals like this one. With radiologist shortages biting across Europe and screening ages extending, the question has shifted from whether AI will carry part of the screening workload to which systems can prove they deserve to.

Ten years ago, autonomous mammography reading was a promise. As of this week, it is a certified medical device — with an autopilot, a flight recorder, and a rule for when the pilot must take over.