Multi-Modal Multi-Agent Reinforcement Learning for Radiology Report Generation: Radiologist-Like Workflow with Clinically Verifiable Rewards

2026-03-19 · Source: cs.CV updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Robotics & Autonomous Systems · Depth: Expert, extended

Summary

MARL-Rad, a novel multi-modal multi-agent reinforcement learning (MARL) framework, has been developed for radiology report generation (RRG). This framework coordinates region-specific agents (left, right, central) and a global integrating agent, optimizing the entire system through clinically verifiable rewards. Unlike prior single-model reinforcement learning or post-hoc agentization, MARL-Rad jointly trains multiple agents in an on-policy manner. Experiments on the MIMIC-CXR and IU X-ray datasets demonstrate that MARL-Rad consistently improves clinical efficacy (CE) metrics such as RadGraph F1, CheXbert F1, and GREEN scores, achieving state-of-the-art performance. The system also enhances laterality consistency and produces more accurate, detail-informed reports, mirroring the workflow of human radiologists.

Key takeaway

For research scientists developing medical AI for diagnostic imaging, MARL-Rad demonstrates that end-to-end optimization of multi-agent systems with clinically verifiable rewards yields superior, more detailed, and clinically accurate radiology reports. You should consider adopting a multi-agent reinforcement learning approach that mirrors clinical workflows to improve the interpretability and performance of your models, especially for tasks requiring fine-grained spatial reasoning.

Key insights

Jointly training multi-agent systems with clinically verifiable rewards significantly enhances radiology report generation accuracy.

Principles

• Workflow-aligned multi-agent optimization improves system performance.
• Clinically verifiable rewards are crucial for medical AI tasks.
• Decomposition into region-specific agents enhances spatial reasoning.

Method

MARL-Rad extends Group Sequence Policy Optimization (GSPO) to a multi-agent setting, using CheXbert accuracy, RadGraph F1, and ROUGE-L as verifiable rewards for end-to-end optimization of region-specific and global integrating agents.

In practice

• Implement region-specific agents for detailed image analysis.
• Utilize CheXbert and RadGraph F1 for clinical reward signals.
• Apply multi-agent RL to other workflow-structured medical AI tasks.

Topics

• Multi-Agent Reinforcement Learning
• Radiology Report Generation
• Chest X-ray Interpretation
• Clinically Verifiable Rewards
• Medical Vision-Language Models

Best for: Research Scientist, AI Researcher, AI Scientist, Machine Learning Engineer

Related on AIssential

• Reinforcement Learning Improves LLM Accuracy and Reasoning in Disease Classification from Radiology Reports — Reinforcement learning with GRPO enhances LLM accuracy and reasoning in medical text classification without explicit reasoning supervision.
• MARCH: Multi-Agent Radiology Clinical Hierarchy for CT Report Generation — A multi-agent framework mimicking radiology hierarchies improves CT report generation fidelity and accuracy.
• SDR: Set-Distance Rewards for Radiology Report Generation — Set-distance rewards unify post-training and test-time scaling for unordered text generation, like radiology reports.
• AgenticRL: Self-Refining Agentic Reinforcement Learning for Vision-Conditioned UAV Navigation — A multimodal GPT agent can autonomously generate, refine, and deploy reinforcement learning policies for UAV navigation tasks.
• UniRG: Scaling medical imaging report generation with multimodal reinforcement learning — Reinforcement learning with clinically meaningful rewards significantly improves medical vision-language model reliability and generality.
• Toward Clinically Acceptable Chest X-ray Report Generation: A Qualitative Retrospective Pilot Study of CXRMate-2 — CXRMate-2 demonstrates a credible pathway to clinically acceptable chest X-ray report generation through multimodal conditioning and reinforcement learning.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.CV updates on arXiv.org.