MARCH: Multi-Agent Radiology Clinical Hierarchy for CT Report Generation

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Radiology

Background:

• Automated 3D radiology report generation faces challenges with clinical hallucinations and lacks iterative verification.
• Current Vision-Language Models (VLMs) function as black-box systems, omitting collaborative oversight crucial in clinical workflows.

Purpose of the Study:

• To introduce MARCH (Multi-Agent Radiology Clinical Hierarchy), a novel multi-agent framework designed to emulate a radiology department's hierarchy for improved report generation.
• To enhance the reliability and clinical fidelity of AI-generated radiology reports by incorporating specialized agent roles and iterative consensus.

Main Methods:

• MARCH employs a hierarchical structure with specialized agents: a Resident Agent for initial drafting using multi-scale CT feature extraction.
• Fellow Agents conduct retrieval-augmented revisions, while an Attending Agent facilitates consensus discourse to resolve discrepancies.
• The framework simulates a collaborative clinical workflow to ensure accuracy and reduce errors.

Main Results:

• MARCH significantly outperformed state-of-the-art baselines on the RadGenome-ChestCT dataset.
• The framework demonstrated superior clinical fidelity and linguistic accuracy in automated radiology report generation.
• Results indicate improved reliability of AI in high-stakes medical domains through hierarchical, multi-agent collaboration.

Conclusions:

• Modeling human-like organizational structures, such as clinical hierarchies, enhances the reliability of AI in medical applications.
• MARCH offers a promising approach to mitigate hallucinations and improve the accuracy of AI-generated radiology reports.
• The multi-agent, collaborative framework represents a significant advancement in clinical AI systems.