STRIVE: Multi-Agent Structured Temporal Reasoning with Integrated Verification for Longitudinal Radiology Report Generation

Junyeong Maeng; Eunsong Kang; Heung-Il Suk

cs.AIAug 25, 2026

STRIVE: Multi-Agent Structured Temporal Reasoning with Integrated Verification for Longitudinal Radiology Report Generation

Authors: Junyeong Maeng, Eunsong Kang, Heung-Il Suk

Organizations: Department of Artificial Intelligence, Korea University, Seoul, Republic of Korea · Graduate School of Data Science, Kangwon National University, Chuncheon, Republic of Korea

Abstract

Longitudinal radiology report generation (LRRG) requires identifying both current findings and their changes relative to a prior study. Existing methods jointly model diagnosis, attribute estimation, temporal comparison, and language generation within implicit representations, which can cause task interference, obscure the evidence underlying each decision, and limit error traceability. They also model progression states as independent labels, ignoring their ordered structure and thus treating missed changes and direction reversals equally. We present STRIVE, Multi-Agent Structured Temporal Reasoning with Integrated Verification for LRRG, which decomposes clinical reasoning into specialized Diagnosis, Attribute, and Temporal Change Agents that produce explicit intermediate evidence. In particular, the Temporal Change Agent is further post-trained using Progression-Aware GRPO, a verifiable, shaped reward that assigns partial credit to direction-preserving errors while scoring direction reversals lowest. STRIVE performs verification at two stages: a deterministic Consistency Gate reconciles the agent outputs before report generation, and a Validation Agent checks whether the generated report is supported by the aggregated clinical evidence. On Longitudinal-MIMIC, STRIVE attains the best clinical efficacy among recent methods and more than doubles Longitudinal Change Concordance (LCC), a measure of temporal agreement with the reference report, over the strongest baseline.

Figures & tables

Figure 1: (a) A longitudinal report includes finding, presence, severity, location, and change. (b) Existing LRRG entangles them in an implicit latent space, generating clinically inconsistent reports. (c) STRIVE decomposes them across specialized agents, forms an explicit clinical state, and verifies the final report.

Figure 2: Overview of STRIVE. Three Clinical Decision Agents infer per-finding presence (Diagnosis), severity and location (Attribute), and change labels (Temporal Change) post-trained with GRPO. A rule-based Consistency Gate resolves diagnosis–change conflicts and aggregates the outputs into a structured clinical state. A Writer Agent then generates the report and a Validation Agent applies state-grounded edits to produce the final report.

Input	Method	Venue	NLG Metrics $\uparrow$						CE Metrics $\uparrow$
Input	Method	Venue	B-1	B-2	B-3	B-4	R-L	MTR	P	R	F1
Single Image	R2Gen	EMNLP’20	0.308	0.190	0.126	0.089	0.266	0.124	0.457	0.290	0.355
	R2GenCMN	ACL’21	0.328	0.203	0.135	0.096	0.273	0.133	0.512	0.359	0.422
	R2GenGPT	Meta-Rad’23	0.396	0.243	0.161	0.108	0.260	0.155	0.495	0.385	0.404
	PromptMRG	AAAI’24	0.384	0.229	0.149	0.104	0.261	0.146	0.517	0.453	0.439
	EKAGen	CVPR’24	0.405	0.245	0.156	0.105	0.264	0.154	0.472	0.404	0.411

Table 1: Comparison with single-image and longitudinal RRG methods on Longitudinal-MIMIC. B-

n

, R-L, and MTR denote BLEU-

n

, ROUGE-L, and METEOR. P, R, and F1 are the micro-averaged CheXbert precision, recall, and F1. Bold and underline indicate the best and second-best result per metric.

Method	LCC $\uparrow$		ReXrank $\uparrow$
Method	LCC-C	LCC-F	1/RadCliQ-v1	BLEU	BERTScore	SembScore	RadGraph-F1	RaTEScore	GREEN
Prefilling	0.089	0.053	0.819	0.176	0.397	0.352	0.180	0.528	0.281
HERGen	0.163	0.100	0.866	0.200	0.410	0.369	0.201	0.538	0.298
STREAM	0.166	0.100	0.820	0.195	0.401	0.333	0.191	0.537	0.277
MLRG	0.141	0.082	0.819	0.193	0.377	0.374	0.189	0.530	0.299
LLM-RG4	0.185	0.115	0.925	0.216	0.428	0.395	0.216	0.554	0.337

Table 2: LCC and ReXrank results on Longitudinal-MIMIC. Bold and underline indicate the best and second-best results per metric. The ablation variant is excluded from the ranking.

Variant	B-1	MTR	CE-F1	LCC-C	LCC-F
Full model	0.466	0.195	0.620	0.394	0.283
w/o Attribute	0.454	0.191	0.615	0.346	0.250
w/o Temporal	0.452	0.192	0.622	0.148	0.095
w/o GRPO	0.464	0.195	0.618	0.389	0.259
w/o Consistency	0.465	0.195	0.620	0.389	0.274
w/o Base draft	0.381	0.171	0.610	0.404	0.280

Table 3: Component ablation on Longitudinal-MIMIC.

Figure 3: Qualitative comparison on Longitudinal-MIMIC. Colors indicate disease states matched to the reference report, while red boxes denote incomplete states. STRIVE captures all findings and their detailed clinical states.

Appendix figures & tables7 assets

Supplementary material from the paper’s appendix.

Appendix

	Train	Val	Test
Patients	$26{,}156$	$203$	$266$
Studies	$92{,}374$	$737$	$2{,}058$
Reports with change statements	$50{,}849$	$414$	$1{,}390$
Finding-level change statements	$117{,}040$	$954$	$3{,}275$

Appendix

Table S1: Longitudinal-MIMIC partitions. Each study is paired with the patient’s most recent prior study, and the data are split at the patient level to prevent patient overlap across partitions.

	Diag.	Attr.	Temporal
	SFT	SFT	SFT	GRPO
Backbone	Gemma	MedGemma	Gemma	SFT ckpt
LoRA $r$ / $\alpha$	16/32	16/32	16/32	16/32
Optimizer	AdamW	AdamW	AdamW	AdamW
Learning rate	$1e^{-4}$	$1e^{-4}$	$1e^{-4}$	$1e^{-5}$
Batch size	$4$	$4$	$4$	$4$

Appendix

Table S2: Optimization settings of the clinical decision agents. GRPO is initialized from the corresponding SFT checkpoint.

Method	Fine change label (F1) $\uparrow$					LCC-F	Direction (F1) $\uparrow$		LCC-C
Method	new	increased	decreased	resolved	stable	LCC-F	worsened	improved	LCC-C
Prefilling	0.000	0.024	0.005	0.000	0.236	0.053	0.027	0.004	0.089
HERGen	0.040	0.071	0.064	0.000	0.325	0.100	0.107	0.057	0.163
STREAM	0.039	0.086	0.039	0.000	0.336	0.100	0.127	0.036	0.166
MLRG	0.000	0.068	0.039	0.000	0.300	0.082	0.077	0.045	0.141
LLM-RG4	0.000	0.057	0.130	0.023	0.368	0.115	0.066	0.121	0.185

Appendix

Table S3: Per-label decomposition of LCC. Each cell is the F1 with which a system recovers the reference changes carrying that label; LCC-F and LCC-C are the macro means over the five fine labels and the three direction classes. The stable column is shared by both levels and is therefore not repeated in the direction block. The last row gives the number of reference changes carrying each label, which is the denominator of the corresponding recall. Bold and underline mark the best and second-best system per column.

Figure S1: Omission rate under the primary direction-only matching protocol, defined as the fraction of reference-stated changes absent from the generated report. STRIVE achieves the lowest omission rate, while removing the Temporal Change Agent increases it to the baseline range.

Figure S2: Additional qualitative comparison on Longitudinal-MIMIC. Colors identify finding-specific statements in the reference and generated reports: cardiomegaly (blue), atelectasis (magenta), pneumothorax (yellow), pleural effusion (green), and support devices (orange).

Method	LCC-C		LCC-F
Method	Value	$95\%$ CI	Value	$95\%$ CI
Prefilling	0.089	[0.079, 0.099]	0.053	[0.046, 0.060]
HERGen	0.163	[0.146, 0.180]	0.100	[0.087, 0.114]
STREAM	0.166	[0.150, 0.184]	0.100	[0.088, 0.113]
MLRG	0.141	[0.126, 0.157]	0.082	[0.072, 0.092]
LLM-RG4	0.185	[0.169, 0.202]	0.115	[0.103, 0.131]

Appendix

Table S4: LCC scores with

95\%

bootstrap percentile confidence intervals computed over the

2{,}058

test study pairs using

B=2{,}000

resamples. STRIVE’s confidence interval does not overlap with that of any baseline at either granularity. Bold and underlined values indicate the best and second-best point estimates, respectively.

Pipeline stage	LCC-C	LCC-F
Raw Temporal Agent outputs	0.342	0.241
after Consistency Gate	0.354	0.250
after Writer, before Validation	0.338	0.233
after Validation	0.394	0.283

Appendix

Table S5: Stage-wise LCC on the same

3{,}275

reference changes. The first two rows score structured change states; the final two score the report before and after Validation.

Radiology Report Generation

Radiology report generation uses medical images and clinical context to produce structured, clinically accurate findings and impressions.

See all 151 papers

TIM: Temporal Decoupling with Iterative Mutual-Refinement Model for Longitudinal Radiology Report Generation
Oct 19, 2026Yiheng Dong, Yi Lin, Shilong Huang +2Radiology Report Generation Radiology VLMs
Huazhong University of Science and Technology, Wuhan, China · HKUST, Hong Kong, China
Standardizing Longitudinal Radiology Report Evaluation via Large Language Model Annotation
Jan 23, 2026Xinyi Wang, Grazziela Figueredo, Ruizhe Li +1Medical Report Generation LLM-Assisted Annotation
School of Computer Science, The University of Nottingham, Nottingham, NG7 2RD, United Kingdom. · School of Medicine, The University of Nottingham, Nottingham, NG7 2RD, United Kingdom.
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

Oct 19, 2026J/K move · Enter open · S save

TIM: Temporal Decoupling with Iterative Mutual-Refinement Model for Longitudinal Radiology Report Generation

Yiheng Dong, Yi Lin, Shilong Huang, Xiyan Yang+1

Huazhong University of Science and Technology, Wuhan, China · HKUST, Hong Kong, China

Radiology Report Generation Radiology VLMs Temporal Reasoning in VLMs

Open paper AskPDF arXiv

Automatic radiology report generation (RRG) aims to translate medical images into diagnostic text, reducing radiologists' workload and standardizing clinical documentation. Nonetheless, existing approaches mainly focus on single-timepoint analysis and fail to capture temporal disease evolution across longitudinal examinations. While recent longitudinal RRG (LRRG) approaches incorporate historical data, they often combine images from different time points within a single representation space, leading to blurred semantics and inconsistent temporal reasoning. In this work, we propose a Temporal Decoupling with Iterative Mutual-Refinement Model (TIM), a two-stage framework that explicitly decouples spatial pathology from temporal progression and iteratively refines reports through mutual feedback. Stage I performs temporal-decoupled representation learning, separating temporal evolution patterns from disease-specific features and generating radiology reports for both prior and current studies. Stage II introduces a mutual report refinement mechanism that identifies diagnostic inconsistencies within prior reports and iteratively rectifies both prior and current reports through error-sensitive feedback. Experiments on the Longitudinal-MIMIC dataset demonstrate that TIM surpasses existing single-image and longitudinal baselines, achieving new state-of-the-art performance across both language and clinical metrics. Code is available at https://github.com/yihengd/TIM.

Abstract

Radiology Report Generation

Explore similar work

TIM: Temporal Decoupling with Iterative Mutual-Refinement Model for Longitudinal Radiology Report Generation

Standardizing Longitudinal Radiology Report Evaluation via Large Language Model Annotation

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