Paper Overview
- Research area: Machine Learning
- Authors: Aubin Charley, Julien Khlaut, Charles Corbière, Pierre Manceron, Corentin Dancette
- Published: 2026-11-06
- arXiv: 2611.09836
- Report generation is simulated on a large multi-modality report corpus.
- The orchestrator is given either:
- (a) a catalog of finding classes consolidated from cleared devices (each available on its cleared modality), or
- (b) an unrestricted catalog where any class can be requested.
- Planning is driven by clinical indications or by a target report.
- Tool outputs are simulated from reference reports, so the comparison isolates tool availability and orchestration, excluding tool error.
- With the cleared catalog and indications only, the orchestrator recovered less than one-third of clinically actionable findings.
- Recovery improved significantly with the unrestricted catalog, and further when planning came from a target report.
- Attributing each missed finding to a cause showed that most were classes not offered by cleared devices on the study's modality; one-third were classes or tools that were available but never invoked.
- Adding the most frequently missing tools closed most of the gap with the unrestricted catalog on held-out reports.
- A smaller orchestrator using the unrestricted catalog forfeited all of the gains.
Summary
Radiology AI has developed along two paths: specialized models that perform well on narrow tasks but cover few clinical findings, and generalist models that address a broader range of tasks but remain prone to hallucination and inaccurate measurements. Combining the two through agentic systems offers a route toward comprehensive radiology reporting, but published systems are tied to one modality or body part.
This paper describes a radiology agentic framework that standardizes how an orchestrator calls modular tools to produce a structured report, and uses it to measure what such a system can do with tools from FDA-cleared radiology AI devices.
Methodology
Key Findings
Conclusion
Tool availability and effective orchestration are two complementary constraints on comprehensive AI-assisted radiology reporting.
---
*Auto-collected on 2026-11-10*