CRC-Router: Risk-Constrained Routing for Medical Agentic AI Systems

What happened

To address the critical risk that autonomous decisions from medical imaging AI can propagate errors to downstream clinical decisions, paired with the lack of reliable triage mechanisms, researchers developed the model-agnostic, risk-constrained routing module CRC-Router. CRC-Router combines multi-dimensional uncertainty signals and prediction scores to build lesion-level routing features, uses a lightweight risk model to estimate the risk of erroneous decision acceptance, and calibrates decision acceptance thresholds aligned with user-specified risk targets via conformal risk control. On chest X-ray multi-finding triage tasks, CRC-Router delivers a better risk-coverage tradeoff than all evaluated baselines, whether deployed as a standalone routing layer or integrated with MedRAX, the state-of-the-art medical AI agent. The module's code is open-source.

Key facts

Module Name

CRC-Router (Risk-Constrained Routing Module)

Use Case

Decision triage for medical imaging prediction models and medical AI agent systems

Test Dataset

NIH ChestX-ray14 chest X-ray multi-finding triage dataset

Core Features

Model-agnostic, deployable as a standalone layer or plug-in to existing medical AI pipelines, compatible with state-of-the-art medical AI agent MedRAX

Code Repository

https://github.com/XLIAaron/CRC-Router

Background

The medical imaging field is increasingly exploring AI agent systems to boost diagnostic throughput and reduce clinical clinicians' workload, but a core barrier to safe, real-world deployment remains: errors from autonomous system judgments can propagate along the clinical workflow to downstream decision points, causing tangible patient harm.

Why it matters

For the medical AI industry, this module fills a key gap for controllable risk triage in selective automation scenarios, pushing medical AI deployment along a path that balances performance with safety and regulatory compliance, rather than focusing solely on raw performance. For developers, its model-agnostic, plug-in design drastically reduces the cost of retrofitting existing medical AI pipelines to meet compliance requirements. For clinical clinicians, the system automatically refers high-risk cases, reducing redundant work while upholding diagnostic safety.