arxiv: 2604.13598 · v1 · submitted 2026-04-15 · 💻 cs.LG · stat.ME

Recognition: unknown

Enhancing Reinforcement Learning for Radiology Report Generation with Evidence-aware Rewards and Self-correcting Preference Learning

Qin Zhou , Guoyan Liang , Qianyi Yang , Jingyuan Chen , Sai Wu , Chang Yao , Zhe Wang

Authors on Pith no claims yet

Pith reviewed 2026-05-10 14:04 UTC · model grok-4.3

classification 💻 cs.LG stat.ME

keywords reinforcement learningradiology report generationevidence-aware rewardspreference learningclinical faithfulnesschest X-rayself-improvementmedical AI

0 comments

The pith

Evidence-aware rewards and self-correcting preference learning produce more clinically faithful radiology reports.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper develops ESC-RL to fix two problems in using reinforcement learning for radiology report generation from images. Current rewards give little guidance on whether reports match specific evidence in the images, and there is no built-in way for the model to improve itself toward clinical standards. GEAR gives rewards at the group level for evidence alignment by boosting correct findings, filling in missed ones, and cutting unsupported ones. SPL builds preference pairs automatically from varied model outputs and has an LLM refine them into better reports without any human labels. Experiments on two standard chest X-ray collections show steady improvements and reach the best known results.

Core claim

The authors claim that combining group-wise evidence-aware alignment rewards with an LLM-driven self-correcting preference learning loop allows reinforcement learning to optimize radiology reports for clinical faithfulness and disease alignment, leading to consistent performance gains on public datasets.

What carries the argument

ESC-RL consisting of the Group-wise Evidence-aware Alignment Reward (GEAR) that scores true positives, false negatives, and false positives separately for evidence grounding, and the Self-correcting Preference Learning (SPL) that synthesizes a preference dataset and refines reports via LLM to enable unsupervised self-improvement.

If this is right

Reports become more grounded in actual image findings through targeted positive and negative feedback.
The system can keep improving during training by using its own outputs to create better preferences.
Clinical faithfulness increases without requiring additional human-annotated data.
State-of-the-art results on chest X-ray report generation suggest readiness for broader testing.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

This method could be adapted to generate reports for other imaging types like CT or MRI if evidence grouping is defined similarly.
Future work might test whether the same rewards reduce specific error types like hallucinated findings in practice.
Integration into clinical systems could lower the rate of AI-generated inaccuracies that require radiologist correction.

Load-bearing premise

The LLM used in SPL produces clinically reliable refined reports and the GEAR scoring correctly measures faithfulness without adding new biases.

What would settle it

A study in which radiologists rate the clinical accuracy of reports from the new method versus baselines and find no improvement, or cases where LLM refinements introduce factual mistakes that affect patient care.

Figures

Figures reproduced from arXiv: 2604.13598 by Chang Yao, Guoyan Liang, Jingyuan Chen, Qianyi Yang, Qin Zhou, Sai Wu, Zhe Wang.

**Figure 2.** Figure 2: Illustration of the Self-correcting Preference [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗

**Figure 3.** Figure 3: Qualitative comparison of reports generated by R2GenRL, PromptMRG, REVTAF, and our method. [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗

**Figure 4.** Figure 4: An example of re-integrating multiple obser [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗

**Figure 5.** Figure 5: Ablation study. (a) The influence of the dif [PITH_FULL_IMAGE:figures/full_fig_p011_5.png] view at source ↗

**Figure 6.** Figure 6: The prompt template used for re-integrating multiple observations into a refined report within the ESC-RL [PITH_FULL_IMAGE:figures/full_fig_p013_6.png] view at source ↗

read the original abstract

Recent reinforcement learning (RL) approaches have advanced radiology report generation (RRG), yet two core limitations persist: (1) report-level rewards offer limited evidence-grounded guidance for clinical faithfulness; and (2) current methods lack an explicit self-improving mechanism to align with clinical preference. We introduce clinically aligned Evidence-aware Self-Correcting Reinforcement Learning (ESC-RL), comprising two key components. First, a Group-wise Evidence-aware Alignment Reward (GEAR) delivers group-wise, evidence-aware feedback. GEAR reinforces consistent grounding for true positives, recovers missed findings for false negatives, and suppresses unsupported content for false positives. Second, a Self-correcting Preference Learning (SPL) strategy automatically constructs a reliable, disease-aware preference dataset from multiple noisy observations and leverages an LLM to synthesize refined reports without human supervision. ESC-RL promotes clinically faithful, disease-aligned reward and supports continual self-improvement during training. Extensive experiments on two public chest X-ray datasets demonstrate consistent gains and state-of-the-art performance.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

The paper pairs group-wise evidence rewards with unsupervised LLM preference synthesis for radiology report RL, but the synthesis step assumes unvalidated LLM outputs are clinically reliable.

read the letter

The paper's main advance is ESC-RL, which adds a Group-wise Evidence-aware Alignment Reward (GEAR) and a Self-correcting Preference Learning (SPL) loop to reinforcement learning for radiology report generation. GEAR scores outputs at the finding-group level to reinforce true positives, recover false negatives, and suppress false positives. SPL builds a preference dataset by feeding multiple noisy model outputs to an LLM that synthesizes a refined report, all without human labels. This directly targets the bluntness of standard report-level rewards and the lack of built-in self-improvement in prior RL-RRG work. The concrete mechanisms are a clear step beyond the cited baselines. The paper does a solid job spelling out those limitations and showing how the two components fit together to produce more evidence-grounded training signals. The unsupervised preference construction is practically motivated for a domain where expert annotation is expensive. The soft spot is SPL itself. The method treats the LLM-refined reports as reliable and disease-aware by construction, yet the description supplies no clinical review, no error-rate comparison against radiologist ground truth, and no ablation confirming the refined versions actually reduce hallucinations or missed findings relative to the noisy inputs. If the LLM introduces its own errors, preference learning will reinforce them. GEAR's gains also depend on the quality of the evidence detector, which receives less scrutiny. The abstract claims consistent gains and SOTA on two chest X-ray datasets, but without the actual tables, baselines, or significance tests visible here, it is impossible to judge whether the improvements are meaningful or mainly from tuning. This is for researchers working on RL for grounded medical text or preference optimization in high-stakes domains. A reader focused on evidence-based generation would get concrete ideas to adapt, even if they would add their own validation layer. It deserves peer review because the components are well-specified and address documented gaps; referees can test the clinical reliability of the synthesized preferences and check the experimental details.

Referee Report

2 major / 2 minor

Summary. The manuscript introduces ESC-RL for radiology report generation, featuring two main components: the Group-wise Evidence-aware Alignment Reward (GEAR) which provides evidence-based feedback to reinforce true positives, recover false negatives, and suppress false positives, and Self-correcting Preference Learning (SPL) which automatically builds a disease-aware preference dataset using an LLM to synthesize refined reports from noisy observations without human supervision. The authors claim that this approach enables clinically faithful rewards and continual self-improvement, leading to consistent performance gains and state-of-the-art results on two public chest X-ray datasets.

Significance. Should the empirical claims be substantiated and the reliability of the LLM-synthesized reports validated, this work could meaningfully advance the application of reinforcement learning in medical imaging by addressing limitations in evidence grounding and self-alignment. It introduces novel mechanisms for preference learning in RRG that may inspire similar self-correcting approaches in other generative tasks.

major comments (2)

[Abstract] The abstract states that 'extensive experiments on two public chest X-ray datasets demonstrate consistent gains and state-of-the-art performance' yet includes no quantitative results, specific metrics, baseline models, ablation studies, or statistical tests. This absence prevents assessment of whether the central claim of superiority is supported by the data.
[SPL component] The Self-correcting Preference Learning (SPL) strategy relies on an LLM to synthesize refined reports treated as reliable ground truth for constructing the preference dataset, without any mentioned clinical validation or comparison to radiologist annotations. This is a load-bearing assumption for the self-improvement claim; if the synthesized reports do not improve upon the noisy inputs in terms of clinical faithfulness, the preference learning may not achieve the intended alignment and could propagate errors.

minor comments (2)

The description of how GEAR computes its group-wise scores and integrates with the base RL algorithm could be expanded for reproducibility.
Ensure that all acronyms (e.g., GEAR, SPL, ESC-RL) are defined at first use and that the experimental setup details, such as the specific RL algorithm used, are clearly stated.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the thoughtful and constructive feedback. We address each major comment below, indicating planned revisions to strengthen the manuscript.

read point-by-point responses

Referee: [Abstract] The abstract states that 'extensive experiments on two public chest X-ray datasets demonstrate consistent gains and state-of-the-art performance' yet includes no quantitative results, specific metrics, baseline models, ablation studies, or statistical tests. This absence prevents assessment of whether the central claim of superiority is supported by the data.

Authors: We agree that the abstract would benefit from including key quantitative highlights to allow immediate assessment of the claims. In the revised version, we will update the abstract to report specific metrics (e.g., BLEU-4, ROUGE-L, and CheXbert F1 improvements), name the main baselines, and note that full ablations and statistical significance tests appear in the experiments section. This change will be made without exceeding typical abstract length constraints. revision: yes
Referee: [SPL component] The Self-correcting Preference Learning (SPL) strategy relies on an LLM to synthesize refined reports treated as reliable ground truth for constructing the preference dataset, without any mentioned clinical validation or comparison to radiologist annotations. This is a load-bearing assumption for the self-improvement claim; if the synthesized reports do not improve upon the noisy inputs in terms of clinical faithfulness, the preference learning may not achieve the intended alignment and could propagate errors.

Authors: We acknowledge this is a substantive concern regarding the core assumption of SPL. The manuscript demonstrates the benefit of SPL through ablation studies showing performance gains and provides examples of synthesized reports, but does not include direct clinical validation against radiologist annotations. In revision, we will expand the SPL section with additional qualitative analysis of report refinements (highlighting evidence grounding improvements) and add an explicit limitations paragraph discussing the potential for error propagation and the value of future radiologist validation. We will also clarify that the LLM synthesis is guided by multiple noisy observations rather than treated as infallible ground truth. revision: partial

Circularity Check

0 steps flagged

No significant circularity; empirical method with independent experimental validation

full rationale

The paper describes an empirical RL framework (ESC-RL) with two components: GEAR for evidence-aware rewards and SPL for constructing preference data via LLM synthesis of refined reports from noisy observations. No equations, derivations, or parameter-fitting steps are shown in the provided text that reduce any claimed prediction or result to its inputs by construction. SPL is presented as a self-correcting mechanism, but its outputs are not mathematically defined in terms of the final performance metric; instead, the paper relies on downstream experiments on public chest X-ray datasets for validation. No self-citations, uniqueness theorems, or ansatzes are invoked in a load-bearing way that collapses the argument. The approach is self-contained against external benchmarks and does not exhibit any of the enumerated circularity patterns.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 2 invented entities

The central claim rests on two unverified assumptions: that LLM-refined reports are clinically trustworthy and that the evidence-aware scoring function correctly measures faithfulness. No free parameters are explicitly named, but reward-component weights and LLM prompting choices are implicit.

free parameters (1)

GEAR component weights
Weights balancing true-positive reinforcement, false-negative recovery, and false-positive suppression must be chosen or tuned.

axioms (1)

domain assumption An off-the-shelf LLM can produce clinically accurate refined reports from noisy model outputs without human oversight
Invoked to construct the preference dataset in SPL.

invented entities (2)

Group-wise Evidence-aware Alignment Reward (GEAR) no independent evidence
purpose: Provide fine-grained, evidence-grounded RL signal for report generation
Newly defined reward that operates on groups of findings rather than whole reports.
Self-correcting Preference Learning (SPL) no independent evidence
purpose: Automatically generate reliable preference pairs for continual RL improvement
New strategy that uses multiple noisy observations plus LLM synthesis.

pith-pipeline@v0.9.0 · 5490 in / 1327 out tokens · 36031 ms · 2026-05-10T14:04:35.524193+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

34 extracted references · 24 canonical work pages · 1 internal anchor

[1]

online" 'onlinestring :=

ENTRY address archivePrefix author booktitle chapter edition editor eid eprint eprinttype howpublished institution journal key month note number organization pages publisher school series title type volume year doi pubmed url lastchecked label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block STRING...
[2]

write newline

" write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...
[3]

Asma Alkhaldi, Raneem Alnajim, Layan Alabdullatef, Rawan Alyahya, Jun Chen, Deyao Zhu, Ahmed Alsinan, and Mohamed Elhoseiny. 2024. https://arxiv.org/abs/2407.04106 Minigpt-med: Large language model as a general interface for radiology diagnosis . Preprint, arXiv:2407.04106

work page arXiv 2024
[4]

Zhihong Chen, Yaling Shen, Yan Song, and Xiang Wan. 2022. https://arxiv.org/abs/2204.13258 Cross-modal memory networks for radiology report generation . Preprint, arXiv:2204.13258

work page arXiv 2022
[5]

Zhihong Chen, Yan Song, Tsung-Hui Chang, and Xiang Wan. 2020. Generating radiology reports via memory-driven transformer. arXiv preprint arXiv:2010.16056

work page arXiv 2020
[6]

InThe Thirty-ninth Annual Conference on Neural Information Processing Systems

Zhihong Chen, Maya Varma, Jean-Benoit Delbrouck, Magdalini Paschali, Louis Blankemeier, Dave Van Veen, Jeya Maria Jose Valanarasu, Alaa Youssef, Joseph Paul Cohen, Eduardo Pontes Reis, Emily B. Tsai, Andrew Johnston, Cameron Olsen, Tanishq Mathew Abraham, Sergios Gatidis, Akshay S Chaudhari, and Curtis Langlotz. 2024. https://arxiv.org/abs/2401.12208 Chex...

work page arXiv 2024
[7]

Jie Cheng, Gang Xiong, Xingyuan Dai, Qinghai Miao, Yisheng Lv, and Fei-Yue Wang. 2024. RIME : Robust preference-based reinforcement learning with noisy preferences. In Proceedings of the 41st International Conference on Machine Learning, volume 235, pages 8229--8247. PMLR

2024
[8]

Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. https://doi.org/10.18653/v1/N19-1423 BERT : Pre-training of deep bidirectional transformers for language understanding . In Proceedings of the 2019 Conference of the North A merican Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long a...

work page doi:10.18653/v1/n19-1423 2019
[9]

Demner Fushman Dina, Marc D Kohli, Marc B Rosenman, Sonya E Shooshan, Rodriguez Laritza, Antani Sameer, George R Thoma, and Clement J Mcdonald. 2015. Preparing a collection of radiology examinations for distribution and retrieval. Journal of the American Medical Informatics Association Jamia, (2):2

2015
[10]

Wenjun Hou, Yi Cheng, Kaishuai Xu, Heng Li, Yan Hu, Wenjie Li, and Jiang Liu. 2025. https://arxiv.org/abs/2505.14318 Radar: Enhancing radiology report generation with supplementary knowledge injection . Preprint, arXiv:2505.14318

work page arXiv 2025
[11]

arXiv preprint arXiv:2106.14463 (2021)

Saahil Jain, Ashwin Agrawal, Adriel Saporta, Steven QH Truong, Du Nguyen Duong, Tan Bui, Pierre Chambon, Yuhao Zhang, Matthew P. Lungren, Andrew Y. Ng, Curtis P. Langlotz, and Pranav Rajpurkar. 2021. https://arxiv.org/abs/2106.14463 Radgraph: Extracting clinical entities and relations from radiology reports . Preprint, arXiv:2106.14463

work page arXiv 2021
[12]

Haibo Jin, Haoxuan Che, Yi Lin, and Hao Chen. 2024. Promptmrg: Diagnosis-driven prompts for medical report generation. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 2607--2615

2024
[13]

Alistair E. W. Johnson, Tom J. Pollard, Nathaniel R. Greenbaum, Matthew P. Lungren, Chih ying Deng, Yifan Peng, Zhiyong Lu, Roger G. Mark, Seth J. Berkowitz, and Steven Horng. 2019. https://arxiv.org/abs/1901.07042 Mimic-cxr-jpg, a large publicly available database of labeled chest radiographs . Preprint, arXiv:1901.07042

work page arXiv 2019
[14]

Elia Kaufmann, Leonard Bauersfeld, Antonio Loquercio, Matthias M \"u ller, Vladlen Koltun, and Davide Scaramuzza. 2023. Champion-level drone racing using deep reinforcement learning. Nature, 620(7976):982--987

2023
[15]

Mingjie Li, Haokun Lin, Liang Qiu, Xiaodan Liang, Ling Chen, Abdulmotaleb Elsaddik, and Xiaojun Chang. 2024. https://arxiv.org/abs/2407.14474 Contrastive learning with counterfactual explanations for radiology report generation . Preprint, arXiv:2407.14474

work page arXiv 2024
[16]

Chin-Yew Lin. 2004. https://aclanthology.org/W04-1013/ ROUGE : A package for automatic evaluation of summaries . In Text Summarization Branches Out, pages 74--81, Barcelona, Spain. Association for Computational Linguistics

2004
[17]

Kang Liu, Zhuoqi Ma, Xiaolu Kang, Yunan Li, Kun Xie, Zhicheng Jiao, and Qiguang Miao. 2025. https://doi.org/10.1109/cvpr52734.2025.00968 Enhanced contrastive learning with multi-view longitudinal data for chest x-ray report generation . In 2025 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), page 10348–10359. IEEE

work page doi:10.1109/cvpr52734.2025.00968 2025
[18]

Sophie Ostmeier, Justin Xu, Zhihong Chen, Maya Varma, Louis Blankemeier, Christian Bluethgen, Arne Edward Michalson Md, Michael Moseley, Curtis Langlotz, Akshay S Chaudhari, and Jean-Benoit Delbrouck. 2024. https://doi.org/10.18653/v1/2024.findings-emnlp.21 Green: Generative radiology report evaluation and error notation . In Findings of the Association f...

work page doi:10.18653/v1/2024.findings-emnlp.21 2024
[19]

Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. 2002. https://doi.org/10.3115/1073083.1073135 B leu: a method for automatic evaluation of machine translation . In Proceedings of the 40th Annual Meeting of the Association for Computational Linguistics, pages 311--318, Philadelphia, Pennsylvania, USA. Association for Computational Linguistics

work page doi:10.3115/1073083.1073135 2002
[20]

Vu Minh Hieu Phan, Yutong Xie, Yuankai Qi, Lingqiao Liu, Liyang Liu, Bowen Zhang, Zhibin Liao, Qi Wu, Minh-Son To, and Johan W Verjans. 2024. Decomposing disease descriptions for enhanced pathology detection: A multi-aspect vision-language pre-training framework. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1...

2024
[21]

Han Qin and Yan Song. 2022. https://doi.org/10.18653/v1/2022.findings-acl.38 Reinforced cross-modal alignment for radiology report generation . In Findings of the Association for Computational Linguistics: ACL 2022, pages 448--458, Dublin, Ireland. Association for Computational Linguistics

work page doi:10.18653/v1/2022.findings-acl.38 2022
[22]

Ng, and Matthew P

Akshay Smit, Saahil Jain, Pranav Rajpurkar, Anuj Pareek, Andrew Y. Ng, and Matthew P. Lungren. 2020. https://arxiv.org/abs/2004.09167 Chexbert: Combining automatic labelers and expert annotations for accurate radiology report labeling using bert . Preprint, arXiv:2004.09167

work page arXiv 2020
[23]

Tim Tanida, Philip Müller, Georgios Kaissis, and Daniel Rueckert. 2023. https://doi.org/10.1109/cvpr52729.2023.00718 Interactive and explainable region-guided radiology report generation . In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), page 7433–7442. IEEE

work page doi:10.1109/cvpr52729.2023.00718 2023
[24]

Chaoyi Wu, Xiaoman Zhang, Ya Zhang, Yanfeng Wang, and Weidi Xie. 2023. Medklip: Medical knowledge enhanced language-image pre-training. Proceedings of the IEEE/CVF International Conference on Computer Vision

2023
[25]

Ting Xiao, Lei Shi, Peng Liu, Zhe Wang, and Chenjia Bai. 2024. https://arxiv.org/abs/2412.08901 Radiology report generation via multi-objective preference optimization . Preprint, arXiv:2412.08901

work page arXiv 2024
[26]

Ting Xiao, Lei Shi, Yang Zhang, HaoFeng Yang, Zhe Wang, and Chenjia Bai. 2025. https://arxiv.org/abs/2505.11983 Online iterative self-alignment for radiology report generation . Preprint, arXiv:2505.11983

work page arXiv 2025
[27]

Wei Xiong, Hanze Dong, Chenlu Ye, Ziqi Wang, Han Zhong, Heng Ji, Nan Jiang, and Tong Zhang. 2024. https://arxiv.org/abs/2312.11456 Iterative preference learning from human feedback: Bridging theory and practice for rlhf under kl-constraint . Preprint, arXiv:2312.11456

work page arXiv 2024
[28]

Heng Yin, Shanlin Zhou, Pandong Wang, Zirui Wu, and Yongtao Hao. 2025. https://aclanthology.org/2025.coling-main.276/ KIA : Knowledge-guided implicit vision-language alignment for chest X -ray report generation . In Proceedings of the 31st International Conference on Computational Linguistics, pages 4096--4108, Abu Dhabi, UAE. Association for Computationa...

2025
[29]

Ng, Curtis P

Feiyang Yu, Mark Endo, Rayan Krishnan, Ian Pan, Andy Tsai, Eduardo Pontes Reis, Eduardo Kaiser Ururahy Nunes Fonseca, Henrique Min Ho Lee, Zahra Shakeri Hossein Abad, Andrew Y. Ng, Curtis P. Langlotz, Vasantha Kumar Venugopal, and Pranav Rajpurkar. 2022. https://doi.org/10.1101/2022.08.30.22279318 Evaluating progress in automatic chest x-ray radiology rep...

work page doi:10.1101/2022.08.30.22279318 2022
[30]

BERTScore: Evaluating Text Generation with BERT

Tianyi Zhang, Varsha Kishore, Felix Wu, Kilian Q. Weinberger, and Yoav Artzi. 2020. https://arxiv.org/abs/1904.09675 Bertscore: Evaluating text generation with bert . Preprint, arXiv:1904.09675

work page internal anchor Pith review arXiv 2020
[31]

Xi Zhang, Zaiqiao Meng, Jake Lever, and Edmond S. L. Ho. 2025. https://arxiv.org/abs/2411.19378 Libra: Leveraging temporal images for biomedical radiology analysis . Preprint, arXiv:2411.19378

work page arXiv 2025
[32]

Topol, and Pranav Rajpurkar

Hong-Yu Zhou, Julián Nicolás Acosta, Subathra Adithan, Suvrankar Datta, Eric J. Topol, and Pranav Rajpurkar. 2025 a . https://arxiv.org/abs/2405.07988 Medversa: A generalist foundation model for medical image interpretation . Preprint, arXiv:2405.07988

work page arXiv 2025
[33]

Qin Zhou, Guoyan Liang, Xindi Li, Jingyuan Chen, Wang Zhe, Chang Yao, and Sai Wu. 2025 b . https://arxiv.org/abs/2507.07568 Learnable retrieval enhanced visual-text alignment and fusion for radiology report generation . Preprint, arXiv:2507.07568

work page arXiv 2025
[34]

Zijian Zhou, Miaojing Shi, Meng Wei, Oluwatosin Alabi, Zijie Yue, and Tom Vercauteren. 2024. https://arxiv.org/abs/2403.06728 Large model driven radiology report generation with clinical quality reinforcement learning . Preprint, arXiv:2403.06728

work page arXiv 2024