A new preprint pitches a traffic cop for medical AI: a module that knows when to answer and when to punt to a human.
Researchers describe CRC-Router, a routing layer meant to sit on top of medical imaging AI and decide, case by case, whether the system's finding is safe to accept or should be escalated for review. It combines several uncertainty signals with the model's predictive score into a per-finding risk feature, feeds that into a lightweight risk model, then uses a statistical technique called Conformal Risk Control to set acceptance thresholds that hold to a target error rate the user picks. According to the arXiv preprint (arXiv:2609.30714, not yet peer-reviewed), the authors tested it on chest X-ray triage using the public NIH ChestX-ray14 dataset and report it beat other baselines on the risk-coverage trade-off, meaning it flagged fewer cases while keeping errors in check. They also say it worked as a plug-in on top of MedRAX, which the paper describes as a state-of-the-art medical AI agent. Code is posted on GitHub at github.com/XLIAaron/CRC-Router.
The pitch matters because agentic AI in radiology is being sold on throughput: more scans read, less clinician time spent. The unsolved problem has always been what happens when the AI is confidently wrong. A calibration layer that bolts onto existing models, rather than one built into a single vendor's pipeline, is the more useful shape for a hospital trying to avoid lock-in.
Worth remembering: this is a self-reported result from the paper's own authors, on one public dataset, paired with one agent. Nobody outside the project has checked the math yet, and chest X-rays are a long way from every imaging modality a hospital runs.