German company Vara has received the CE mark for an artificial intelligence system capable of independently triaging mammography examinations, in what the company described as the first European approval for a system that can issue results for examinations it classifies as normal without review by a radiologist. The approval comes amid a shortage of radiologists working in breast cancer screening programs.
The system does not eliminate human review for all examinations. Every examination that the artificial intelligence classifies as abnormal or requiring additional evaluation must be reviewed by at least one radiologist. Examinations that the system identifies as normal can have their reports issued automatically through the independent-triage feature.
Medical Approval with Continuous Oversight
The system was approved as a Class IIb medical device under the European Union’s Medical Device Regulation, a class indicating a medium to high level of risk. Vara said that the safety system it developed to oversee artificial intelligence models helped it meet the requirement for continuous human oversight of high-risk artificial intelligence systems under the European AI Act.
The company calls this system ATMON, short for Autonomous-Triage Monitoring. It monitors changes in breast imaging equipment, the system’s status, and daily performance. If these indicators move outside predefined limits, the system can return the site to full reading by radiologists.
What Changes in Practice?
Vara says the independent-triage feature is now available to population screening programs across Europe, with rollout taking place according to each country’s guidelines and the readiness of its programs. In Germany, more than half of organized breast screening programs operate on Vara systems, processing more than 250,000 examinations per month, according to the company.
Vara also makes the ATMON system available to new and existing customers even if they do not choose to transition to fully independent triage. This means the monitoring mechanism can be used as an additional safety layer in scenarios relying on human reading or on artificial intelligence support alone.
Why Does This Matter?
The significance here is not merely adding an artificial intelligence model to the radiology workflow, but allowing it to close out a portion of normal examinations without a physician’s reading, while keeping other cases within human review. This makes performance monitoring during operation a central factor, because model accuracy may be affected by changes in imaging equipment or usage conditions compared with the study environment.
The company cited a large study it funded and published in Nature Medicine last year. The study, conducted between 2021 and 2023 and involving more than 460,000 women, compared traditional double reading by radiologists with artificial-intelligence-supported double reading and found a higher breast cancer detection rate in the artificial-intelligence-supported readings. However, approval of independent triage raises a different practical question from the model’s performance in support: How effective are monitoring mechanisms at detecting performance deterioration before it affects screening results?