MARCH Framework Uses Multi-Agent Hierarchy to Improve Automated Radiology Report Generation

A newly created multi-agent system known as MARCH (Multi-Agent Radiology Clinical Hierarchy) aims to overcome challenges in the automated generation of 3D radiology reports. This framework mimics the hierarchical structure of radiology departments by designating specific roles to various agents. Traditional Vision-Language Models (VLMs) typically function as isolated "black-box" entities, lacking the collaborative oversight inherent in clinical settings, which can lead to clinical hallucinations and inadequate iterative checks. MARCH incorporates a Resident Agent for initial report drafting through multi-scale CT feature extraction, several Fellow Agents for retrieval-augmented revisions, and an Attending Agent to facilitate a consensus discussion for resolving diagnostic inconsistencies. According to arXiv preprint 2604.16175v1, MARCH significantly surpasses existing state-of-the-art benchmarks on the RadGenome-ChestCT dataset, striving to integrate human-like iterative verification into automated processes.