PACER: Progressive Availability-Conditioned Evidence Routing for Radiology Report Generation under Incomplete Clinical Context

Yulong Chen; Yadong Liu; Haoyu Cao; Sen Xu; Yueying Wang; Jie Wen

cs.CVSep 28, 2026

PACER: Progressive Availability-Conditioned Evidence Routing for Radiology Report Generation under Incomplete Clinical Context

Authors: Yulong Chen, Yadong Liu, Haoyu Cao, Sen Xu, Yueying Wang, Jie Wen

Organizations: Harbin Institute of Technology, Shenzhen, China

Abstract

Radiology report generation (RRG) increasingly incorporates heterogeneous clinical evidence, such as multi-view radiographs and previous reports, whose availability varies across examinations. However, accommodating different input combinations does not ensure effective evidence use: generated reports may still omit or inaccurately describe clinically relevant findings. To address this problem, we propose PACER, a Progressive Availability-Conditioned Evidence Routing framework for structured incomplete-context RRG that follows a Refine-Calibrate-Commit pipeline. It first refines observed visual representations through endpoint-preserving patchwise routing across frozen encoder depths, incorporating complementary cues while retaining the pretrained terminal representation. It then calibrates the language-model prefix according to the observed evidence and availability state, adapting the shared generator's conditioning as the available source set changes. Finally, it generates polarity-structured clinical commitments before the report in the same autoregressive trajectory, providing structured clinical context for subsequent generation. Experiments demonstrate state-of-the-art clinical efficacy across all four MIMIC-RG4 settings and strong MIMIC-CXR performance, while maintaining competitive language-generation quality.

Figures & tables

Figure 1: Overview of Pacer under structured incomplete clinical context. Refine enhances observed visual evidence through endpoint-preserving patchwise depth routing, Calibrate applies availability-conditioned prefix calibration, and Commit generates polarity-structured clinical commitments before the report in the same autoregressive trajectory.

Setting	Model	CE Metrics			NLG Metrics
Setting	Model	P	R	F1	B@1	B@2	B@3	B@4	R-L	MTR
SN	CXRMate ∗	0.572	0.560	0.566	0.421	0.271	0.179	0.122	0.311	0.174
	RadFM	0.413	0.303	0.350	0.188	0.090	0.048	0.028	0.190	0.094
	LLM-RG4	0.588	0.632	0.609	0.479	0.343	0.255	0.196	0.384	0.209
	SimMLM †	0.568	0.591	0.579	0.425	0.277	0.186	0.128	0.313	0.168
	RAGPT †	0.565	0.564	0.564	0.423	0.278	0.188	0.129	0.314	0.168

Table 1: Four-context results on MIMIC-RG4. Published baselines follow Wang et al. (2025b) ;

\ast

denotes CXRMate retrained therein and

\dagger

our SimMLM/RAGPT adaptations. Bold/underline indicate the best/second-best non-tied values per context; ties for best are bold.

Literature-reported results under original protocols
Model	CE Metrics			Clean NLG			Original NLG
Model	P	R	F1	B@1	B@4	R-L	B@1	B@4	R-L
KiUT ( Huang et al., 2023 )	0.371	0.318	0.321	–	–	–	0.393	0.113	0.285
RGRG ( Tanida et al., 2023 )	0.461	0.475	0.447	–	–	–	0.373	0.126	0.264
EKAGen ( Bu et al., 2024 )	0.517	0.483	0.499	–	–	–	0.419	0.119	0.287
MAIRA-1 (7B) ( Hyland et al., 2023 )	–	–	0.553	–	–	–	0.392	0.142	0.289

Table 2: Conventional findings-only SN results. The upper block provides literature context under original protocols, while the lower block evaluates methods on a common findings-only SN test set, with clean and original references reported separately.

\dagger

denotes local downstream training initialized from released pretrained weights, and

\ddagger

denotes a released checkpoint evaluated by us. Bold/underline indicate best/second-best values within the common-test block.

Model	P	R	F1	B@1	B@4	R-L
Base	0.576	0.587	0.581	0.461	0.198	0.395
Refine only	0.575	0.598	0.586	0.470	0.200	0.396
Commit only	0.599	0.587	0.593	0.455	0.194	0.399
Refine + Commit	0.598	0.603	0.600	0.452	0.192	0.396
PACER (+ Calibrate)	0.603	0.617	0.610	0.466	0.198	0.399

Table 3: Component ablations averaged over the four MIMIC-RG4 contexts. Base denotes the local ablation baseline. Bold/underline indicate the best/second-best values.

Mechanism	Variant	P	R	F1	B@1	B@4	R-L
Refine	No Refine (endpoint only)	0.599	0.587	0.593	0.455	0.194	0.399
	Global depth weighting	0.588	0.603	0.595	0.444	0.185	0.390
	Patchwise depth routing	0.598	0.603	0.600	0.452	0.192	0.396
Commit	No Commit (direct report)	0.575	0.598	0.586	0.470	0.200	0.396
	Always commitment-first	0.594	0.596	0.595	0.448	0.190	0.394
	Stochastic trajectory routing	0.598	0.603	0.600	0.452	0.192	0.396

Table 4: Mechanism ablations averaged over the four MIMIC-RG4 contexts. For prefix routing,

h

z

, and

e

denote the prefix summary, observed-evidence summary, and availability embedding. Bold/underline indicate best/second-best values within each group.

Figure 2: Qualitative comparison of longitudinal evidence utilization, showing how the generated report changes when the previous report becomes available.

Appendix figures & tables3 assets

Supplementary material from the paper’s appendix.

Appendix

Phase	Context	LR	Batch	Accum.	Updates
Refine warm-up	SN	$3{\times}10^{-4}$	24	2	14,384
Refine+Commit parent	SN/SW/MN/MW	$3{\times}10^{-4}$	16	2	43,152
Calibrate	SN/SW/MN/MW	$1{\times}10^{-4}$	16	2	21,576

Appendix

Table 5: Optimization protocol used for PACER. “Batch” denotes per-device batch size and “Accum.” gradient accumulation.

Context	Train	Validation	Test
SN	172,608	1,391	2,357
SW	112,776	937	2,026
MN	91,341	701	1,004
MW	47,686	371	828

Appendix

Table 6: Scenario-record counts in the four MIMIC-RG4 availability states.

Method	P	R	F1	B@1	B@4	R-L
MLRG checkpoint	0.536	0.402	0.460	0.365	0.097	0.269
RGRG (official beam4)	0.515	0.523	0.519	0.366	0.105	0.264
RGRG (greedy)	0.489	0.603	0.540	0.362	0.109	0.271
RGRG (dedup.)	0.517	0.515	0.516	0.294	0.093	0.254
EKAGen (missing $\rightarrow$ empty)	0.528	0.442	0.481	0.346	0.089	0.275
EKAGen (nearest-neighbor)	0.529	0.444	0.483	0.346	0.089	0.276

Appendix

Table 7: Supplementary checkpoint-transfer and common-test SN evaluations. RGRG decoding settings and EKAGen missing-entry handling are protocol variants rather than independent models.

Explore similar work

CogRad: A Cognitively-Inspired Multi-Agent Framework for Radiology Report Generation
Jul 4, 2026Saif Ur Rehman Khan, Hasaan Maqsood, Sebastian Vollmer +2Radiology Report Generation Report
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
CCS: Clinical Consensus Selection for Radiology Report Generation
May 28, 2026Xi Zhang, Yingshu Li, Zaiqiao Meng +2Radiology Report Generation Report
School of Computing Science, University of Glasgow · School of Electrical and Computer Engineering, University of Sydney · Language Technology Lab, University of Cambridge
PDD-RRG: Posterior Diagnostic Decision for Study-level Radiology Report Generation
Aug 4, 2026Yang Yu, Yiming Ji, Bin Dai +4Radiology Report Generation Chest X-Ray
School of Computer Science & Technology, Soochow University · Suzhou Institute of Biomedical Engineering and Technology · Jiangsu Key Lab of Language Computing, Suzhou

Jul 4, 2026, cs.CVJ/K move · Enter open · S save

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

Saif Ur Rehman Khan, Hasaan Maqsood, Sebastian Vollmer, Andreas Dengel+1

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

Radiology Report Generation Report Model-Based Multi-Agent Systems

Open paper AskPDF arXiv

Automated radiology report generation (RRG) can ease radiologist workload, yet most existing systems produce a report in a single forward pass, with no mechanism to check a claim against the image or revisit a finding once stated. We present CogRad, a cognitively inspired multi-agent framework that structures generation around four stages of a radiologist's reading process. A Scout agent discovers anatomical regions directly from image patches via slot attention and assigns region and disease-level triage scores; an Investigator agent concentrates representational capacity on the regions Scout flags as suspicious; a Writer agent compiles these signals into a disease gated visual prefix for a large language model; and a Verifier agent supervises training with a visual entailment loss and, at inference, re-examines its own draft sentence by sentence, regenerating any report it judges insufficiently grounded. On CheXpert Plus, CogRad attains a BLEU-4 of 0.316 and a CIDEr of 0.322, the best scores among the methods we compare against. On IU X-Ray, it attains a BLEU-4 of 0.201 and a CIDEr of 0.724, leading every baseline on every standard NLG metric. We further evaluate CogRad with RadGraph F1, CheXbert F1, and a hallucination analysis to assess clinical accuracy beyond standard text-overlap metrics, complemented by ablation studies and Grad-CAM-based visualizations that characterize each agent's contribution and the model's visual grounding.

Abstract

Explore similar work

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

CCS: Clinical Consensus Selection for Radiology Report Generation

PDD-RRG: Posterior Diagnostic Decision for Study-level Radiology Report Generation