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

Kaito Baba; Risa Kishikawa; Satoshi Kodera

cs.CVFeb 17, 2026

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

Authors: Kaito Baba, Risa Kishikawa, Satoshi Kodera

Organizations: Department of Cardiovascular Medicine The University of Tokyo Hospital, Tokyo, Japan

Abstract

We propose MARL-Rad, a multi-modal multi-agent reinforcement learning framework for radiology report generation that trains the entire agentic system on policy within its deployed radiology workflow. MARL-Rad addresses the limitation of post-hoc agentization, where fixed LLMs are organized into hand-designed agentic workflows without being optimized for their assigned roles. Our framework decomposes chest X-ray interpretation into region-specific agents and a global integrating agent, and jointly optimizes them using clinically verifiable rewards. Experiments on the MIMIC-CXR and IU X-ray datasets show that MARL-Rad consistently improves clinical efficacy metrics such as RadGraph, CheXbert, and GREEN scores, achieving state-of-the-art clinical efficacy performance. Further analyses show that MARL-Rad improves laterality consistency and produces more accurate and detailed reports. A blinded clinician evaluation further suggests that MARL-Rad produces reports clinically comparable to ground-truth reports.

Figures & tables

Figure 1: Overview of the proposed multi-agent RL framework. Region-specific agents and the global integrating agent collaboratively generate the radiology report, and the entire agent system is jointly optimized through RL based on clinically verifiable rewards.

Method	NLG Metrics $\ \uparrow$				CE Metrics $\ \uparrow$
Method	BLEU-1	BLEU-4	METEOR	ROUGE-L	RadGraph F1	CheXbert F1	GREEN
MIMIC-CXR Findings
R2Gen [ 16 ]	0.353	0.103	0.142	0.277	-	0.276	-
R2GenCMN [ 15 ]	0.353	0.106	0.142	0.278	-	0.278	-
PPKED [ 68 ]	0.360	0.106	0.149	0.284	-	-	-
CMCL [ 67 ]	0.344	0.097	0.133	0.281	-	-	-

Table 1: Comparison with previous state-of-the-art methods on the MIMIC-CXR dataset [ 48 ] . The best score for each metric within each section is highlighted in bold . Our approach achieves the best performance across all clinical efficacy (CE) metrics, which are considered to align more closely with clinicians’ assessments [ 105 , 69 , 9 , 129 ] .

Method	NLG Metrics $\ \uparrow$				CE Metrics $\ \uparrow$
Method	BLEU-1	BLEU-4	METEOR	ROUGE-L	RadGraph F1	CheXbert F1	GREEN
IU X-ray Findings
R2Gen [ 16 ]	0.470	0.165	0.187	0.371	-	-	-
R2GenCMN [ 15 ]	0.475	0.170	0.191	0.376	-	-	-
PPKED [ 68 ]	0.483	0.168	0.190	0.376	-	-	-
METransformer [ 113 ]	0.483	0.172	0.192	0.380	-	-	-

Table 2: Comparison with previous state-of-the-art methods on the IU X-ray dataset [ 25 ] . The best score for each metric within each section is highlighted in bold . Our approach achieves the best performance across all clinical efficacy (CE) metrics, which are considered to align more closely with clinicians’ assessments [ 105 , 69 , 9 , 129 ] .

Method	NLG Metrics $\ \uparrow$				CE Metrics $\ \uparrow$
Method	BLEU-1	BLEU-4	METEOR	ROUGE-L	RadGraph F1	CheXbert F1	GREEN
MIMIC-CXR Findings + Impression
MedGemma (vanilla)	0.333	0.052	0.338	0.197	0.209	0.523	0.388
MedGemma (agent)	0.252	0.044	0.267	0.144	0.106	0.309	0.344
MedGemma + RL	0.465	0.116	0.373	0.283	0.299	0.502	0.377
Counterfactual Reward	0.575	0.136	0.375	0.272	0.292	0.518	0.369

Table 3: Ablation study results. “MedGemma (vanilla)” represents the unmodified model without reinforcement learning or agentization. “MedGemma (agent)” employs the same agentic workflow as MARL-Rad but without any RL optimization. “MedGemma + RL” applies GSPO to MedGemma in a single-agent setting. “Counterfactual Reward” uses agent-specific rewards based on counterfactual removal of each regional agent’s output from the global agent’s context. The best and worst scores for each metric within each dataset are highlighted in bold and underline, respectively.

Method	Left	Right	Left lung	Right lung
MIMIC-CXR Findings + Impression
MedGemma w/ RL	0.144	0.142	0.219	0.102
MARL-Rad (Ours)	0.184	0.277	0.406	0.146
Left Agent	0.214	0.011	0.369	0.066
Right Agent	0.001	0.257	0.000	0.156
IU X-ray Findings + Impression

Table 4: Laterality-specific RadGraph F1 scores. “Left Agent” and “Right Agent” denote the pre-integration outputs of the corresponding regional agents. Higher scores indicate better laterality consistency. The best score among the final-report methods, i.e., MedGemma w/ RL and MARL-Rad (Ours), is highlighted in bold .

Figure 2: Example output from MARL-Rad. Region-specific agents focus on their assigned regions and generate regional diagnoses. The global integrating agent synthesizes these drafts into a concise and coherent report while adding relevant global findings. The resulting final report is detailed and aligns well with the ground truth, correctly concluding with “No acute cardiopulmonary process.”

Panel of Figure 3 — Figure 3: An example case generated by the single-agent RL model (MedGemma w/ RL) using the same input image as in Fig. 2 .

Figure 4: Physician preferences comparing MARL-Rad and the ground-truth reports. Bars show 5-point Likert ratings for five criteria; the rightmost panel shows the overall 2-choice vote. During evaluation, report sources were anonymized and randomly shuffled for each case.

Explore similar work

MAC-RRG: Iterative Multi-Agent Collaboration for X-ray Radiology Report Generation
Sep 20, 2026Futian Wang, Yuhan Qiao, Xiao Wang +5Radiology Report Generation Multi-Agent Collaboration
School of Computer Science and Technology, Anhui University, Hefei 230601, China · The First Affiliated Hospital of Anhui Medical University, Hefei 230000, China · Harbin Institute of Technology, Shenzhen, China +1
CogRad: A Cognitively-Inspired Multi-Agent Framework for Radiology Report Generation
Jul 4, 2026Saif Ur Rehman Khan, Hasaan Maqsood, Sebastian Vollmer +2Radiology Report Generation Medical VLMs
Department of Computer Science, Rhineland-Palatinate Technical University of Kaiserslautern-Landau, Kaiserslautern, 67663, Germany. · 2German Research Center for Artificial Intelligence, Kaiserslautern, 67663, Germany. · 3Intelligentx GmbH (intelligentx.com), Kaiserslautern, Germany. +1
MARCH: Multi-Agent Radiology Clinical Hierarchy for CT Report Generation
Apr 17, 2026Yi Lin, Yihao Ding, Yonghui Wu +1Radiology Report Generation Multi-Agent Collaboration
Weill Cornell Medicine, New York, USA · University of Western Australia, Crawley, Australia · University of Florida, Florida, USA

Sep 20, 2026J/K move · Enter open · S save

MAC-RRG: Iterative Multi-Agent Collaboration for X-ray Radiology Report Generation

Futian Wang, Yuhan Qiao, Xiao Wang, Dan Xu+4

School of Computer Science and Technology, Anhui University, Hefei 230601, China · The First Affiliated Hospital of Anhui Medical University, Hefei 230000, China · Harbin Institute of Technology, Shenzhen, China+1

Radiology Report Generation Multi-Agent Collaboration

Open paper AskPDF arXiv

Despite the remarkable progress of LLM-based and knowledge graph-augmented Radiology Report Generation (RRG) methods, existing techniques still suffer from inherent defects. Conventional LLM-only models lack structured medical prior knowledge, resulting in frequent medical hallucinations and low diagnostic interpretability. Current knowledge graph-enhanced schemes adopt static one-round knowledge fusion with single-source knowledge, incapable of dynamic knowledge updating according to generation feedback. This paper proposes a novel Multi-Agent Collaborative iterative framework for X-ray Radiology Report Generation, termed MAC-RRG. Inspired by multi-agent technology, our framework constructs a closed-loop optimization paradigm based on task decoupling and collaborative reasoning. Specifically, the framework first generates a preliminary radiology report from input X-ray images via a vision encoder and a basic LLM. Subsequently, a multimodal knowledge graph (MM-KG) agent mines structured disease correlation and anatomical knowledge from medical knowledge graphs, while an auxiliary knowledge agent extracts unstructured domain knowledge from public medical databases. The multi-source knowledge acquired by dual agents is fused and embedded to guide the LLM in iteratively refining the initial report. Extensive quantitative and qualitative experiments on mainstream X-ray RRG datasets, including IU X-ray, MIMIC, and CheXpert Plus, fully verify the superiority of our proposed method. The source code and pre-trained models have been released on https://github.com/Event-AHU/Medical_Image_Analysis

Abstract

Explore similar work

MAC-RRG: Iterative Multi-Agent Collaboration for X-ray Radiology Report Generation

CogRad: A Cognitively-Inspired Multi-Agent Framework for Radiology Report Generation

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