Multi-Modal Multi-Agent Reinforcement Learning for Radiology Report Generation

Researchers introduce MARL-Rad, a multi-agent reinforcement learning framework that optimizes AI agents specifically for radiology report generation rather than using fixed LLMs in pre-designed workflows. The system decomposes chest X-ray interpretation into specialized regional agents coordinated by a global integrator, achieving state-of-the-art clinical performance on benchmark datasets with clinician validation.

MARL-Rad addresses a fundamental limitation in current AI system design: the practice of taking pre-trained language models and arranging them into agentic workflows without optimizing them for their specific roles. This research demonstrates that joint training of multi-agent systems on task-specific objectives produces meaningfully better clinical outcomes than post-hoc organizational approaches.

The framework's decomposition strategy—assigning region-specific agents to different areas of chest X-rays while employing a global integrating agent—mirrors how radiologists actually process images systematically. By training these agents end-to-end using clinically verifiable rewards (RadGraph, CheXbert, GREEN scores), the system learns to generate reports that better align with clinical standards rather than optimizing for generic language quality metrics.

This work carries significant implications for AI-assisted medical diagnostics and other specialized domains requiring nuanced decision-making. Healthcare institutions seeking AI augmentation need systems that achieve clinical efficacy, not just technical performance. The blinded clinician evaluation showing reports comparable to ground-truth benchmarks suggests the framework is approaching practical deployment viability. The improvements in laterality consistency and report accuracy demonstrate that specialized agent training captures domain-specific constraints that generalist models miss.

Looking forward, this research pattern—domain-optimized multi-agent systems rather than orchestrated general-purpose models—likely becomes standard for high-stakes applications. Other medical imaging specialties, legal document analysis, and financial reporting could benefit from similar agent-specialization approaches. The methodology validates that agentic AI requires purpose-built optimization, not generic LLM arrangement.

• →Multi-agent systems optimized jointly for specific roles outperform fixed LLMs arranged in pre-designed workflows
• →Clinical performance metrics improve significantly when training incorporates domain-specific rewards like RadGraph and CheXbert
• →Specialized regional agents with global integration better replicate radiologist reasoning patterns than monolithic approaches
• →Clinician evaluation shows MARL-Rad reports achieve clinical quality comparable to expert radiologist benchmarks
• →Task-specific agent decomposition addresses the broader limitation of post-hoc agentization in AI system design