MARCH: Multi-Agent Radiology Clinical Hierarchy for CT Report Generation
Summary: arXiv:2604.16175v1 Announce Type: new
Abstract: Automated 3D radiology report generation often suffers from clinical hallucinations and a lack of the iterative verification found in human practice. While recent Vision-Language Models (VLMs) have advanced the field, they typically operate as monolithic “black-box” systems without the collaborative oversight characteristic of clinical workflows. To address these challenges, we propose MARCH (Multi-Agent Radiology Clinical Hierarchy), a multi-agent framework that emulates the professional hierarchy of radiology departments and assigns specialized roles to distinct agents.
MARCH utilizes a structured approach to enhance the accuracy and reliability of automated CT report generation. The framework comprises several key components:
- Resident Agent: Responsible for the initial drafting of reports, this agent employs multi-scale CT feature extraction to generate the first version of the report.
- Fellow Agents: These agents engage in retrieval-augmented revision, ensuring that the initial draft is refined based on relevant case studies and existing literature.
- Attending Agent: This senior agent orchestrates an iterative, stance-based consensus discourse, allowing for the resolution of any diagnostic discrepancies that may arise during the report generation process.
One of the significant advancements of MARCH is its ability to simulate real-world clinical workflows, which often involve multiple levels of review and collaborative input. This structured hierarchy not only mirrors the traditional radiology department’s organization but also enhances the functioning of AI in high-stakes medical environments.
In testing, MARCH was evaluated on the RadGenome-ChestCT dataset, where it demonstrated substantial improvements over existing state-of-the-art baselines. The results indicated that MARCH achieved higher clinical fidelity and linguistic accuracy, marking a significant step forward in the field of automated radiology.
The implications of MARCH are profound for the future of medical imaging and AI integration. By addressing the common pitfalls of automated report generation—such as clinical hallucinations and lack of iterative validation—this framework paves the way for more reliable and clinically relevant AI applications. It emphasizes the importance of modeling human-like organizational structures to enhance the reliability of AI systems in critical medical domains.
As AI continues to evolve, frameworks like MARCH will play a crucial role in ensuring that automated systems can operate safely and effectively within the healthcare ecosystem. The success of this multi-agent approach could inspire further research into collaborative AI systems that enhance the capabilities of healthcare professionals while maintaining high standards of clinical excellence.
In conclusion, MARCH represents a significant innovation in the realm of radiology, showcasing the potential for AI to collaborate in ways that mirror human expertise and decision-making. The future of automated radiology reporting looks promising, with frameworks like MARCH leading the charge towards improved accuracy, reliability, and clinical relevance.