MARCH multi-agent framework improves CT radiology report generation
Researchers Yi Lin, Yihao Ding, and Yonghui Wu propose MARCH, a multi-agent framework that mirrors radiology department hierarchies to reduce clinical hallucinations in automated 3D CT report generation.
Score breakdown
Agentic framework designers can draw on MARCH's role-differentiated, hierarchy-mirroring architecture as a blueprint for reducing hallucinations in other high-stakes, multi-step AI reasoning tasks.
- 01MARCH stands for Multi-Agent Radiology Clinical Hierarchy, proposed by Yi Lin, Yihao Ding, and Yonghui Wu.
- 02The framework targets clinical hallucinations and lack of iterative verification in automated 3D CT report generation.
- 03A Resident Agent handles initial drafting using multi-scale CT feature extraction.
Yi Lin, Yihao Ding, and Yonghui Wu introduce MARCH (Multi-Agent Radiology Clinical Hierarchy), a multi-agent framework designed to overcome two core weaknesses in automated 3D CT report generation: clinical hallucinations and the absence of the iterative verification that characterizes real-world radiology practice. Existing Vision-Language Models (VLMs) typically function as monolithic "black-box" systems, lacking the collaborative oversight present in clinical workflows. MARCH addresses this by emulating the professional hierarchy of a radiology department and assigning each tier a distinct, specialized agent role.
A Resident Agent produces an initial draft using multi-scale CT feature extraction.
The framework consists of three agent types. A Resident Agent produces an initial draft using multi-scale CT feature extraction. Multiple Fellow Agents then perform retrieval-augmented revision of that draft. Finally, an Attending Agent oversees an iterative, stance-based consensus discourse among the agents to resolve any remaining diagnostic discrepancies. This layered structure mirrors how junior residents, fellows, and attending physicians collaborate in practice.
Tested on the RadGenome-ChestCT dataset, MARCH significantly outperforms state-of-the-art baselines on both clinical fidelity and linguistic accuracy metrics. The authors argue that modeling human-like organizational structures enhances the reliability of AI in high-stakes medical domains, suggesting that agentic, role-differentiated architectures may be broadly applicable beyond radiology.
Key facts
- 01MARCH stands for Multi-Agent Radiology Clinical Hierarchy, proposed by Yi Lin, Yihao Ding, and Yonghui Wu.
- 02The framework targets clinical hallucinations and lack of iterative verification in automated 3D CT report generation.
- 03A Resident Agent handles initial drafting using multi-scale CT feature extraction.
- 04Multiple Fellow Agents perform retrieval-augmented revision of the initial draft.
- 05An Attending Agent orchestrates iterative, stance-based consensus discourse to resolve diagnostic discrepancies.
- 06MARCH is evaluated on the RadGenome-ChestCT dataset, significantly outperforming state-of-the-art baselines in clinical fidelity and linguistic accuracy.
- 07The framework emulates the professional hierarchy of radiology departments rather than using a monolithic VLM.