List of Papers Browse by Subject Areas Author List
Abstract
Automated radiology report generation (RRG) has gained increasing attention because it can reduce the heavy workload of clinical report writing. However, most existing methods mainly optimize for natural language generation (NLG) metrics that focus on language fluency, while providing little control over clinically important factors such as precision and recall. As consequence, generated reports may be fluent but not well aligned with different clinical needs. To address this challenge, we propose a reinforcement learning framework for precision recall controllable RRG, where a control parameter explicitly adjusts the trade-off between clinical precision and recall during inference. This design allows the model to flexibly generate reports according to different clinical requirements. To ensure clinical correctness, we introduce a clinical reward into the training objective, which helps improve clinical efficacy (CE) beyond standard language based optimization. In addition, we apply a group-relative training strategy that normalizes rewards within each training group, reducing reward variance and improving training stability. Extensive experiments on the MIMIC-CXR dataset show that our method consistently outperforms state-of-the-art approaches in both NLG and CE evaluation metrics, while providing reliable control over the CE precision recall trade-off.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/4076_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/98lingchen/MICCAI2026/
Link to the Dataset(s)
MIMIC-CXR: https://physionet.org/content/mimic-cxr-jpg/2.0.0/
BibTex
@InProceedings{CheLin_Precision_MICCAI2026,
author = { Chen, Ling AND Jin, Ruinan AND Luo, Jun AND Chen, Hanliang AND Strotzer, Quirin AND Yan, Rongkai AND Xue, Yuan AND Prevedello, Luciano AND Wu, Dufan},
title = { { Precision Recall Controllable Radiology Report Generation via Hybrid Natural Language and Clinical Reward Learning } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16878},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
The paper proposes a reinforcement learning framework for radiology report generation that enables explicit control over the precision-recall trade-off of clinically relevant findings. This is achieved through a continuous control parameter that jointly conditions the decoder representations via a conditioned normalization module (PRC-AdaLN) and the clinical reward used during training. The method combines language-based and clinical efficacy rewards and incorporates group-relative advantage normalization to stabilize optimization. Experiments on MIMIC-CXR dataset show improvements in both standard NLG metrics and CheXbert-based clinical metrics.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
- The paper proposes a coherent and practically relevant framework for precision-recall controllable report generation, unifying representation-level conditioning and reward-level shaping through a single control parameter λ.
- The method integrates reinforcement learning with a hybrid clinical and language reward, and uses group-relative normalization to stabilize training, showing consistent improvements over prior work across both NLG and clinical efficacy metrics.
- The experimental section is reasonably good, including multi-metric evaluation, ablations, and analysis of the precision-recall trade-off, and the overall framework is clearly presented and easy to follow.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
- The “clinical loss” is mentioned as an added training objective and ablated, but it is never formally defined as a separate loss term; instead, only a CE-based reward is introduced within the RL objective. This terminology inconsistency creates ambiguity about what is actually optimized beyond CE-derived rewards.
- The implementation of PRC-AdaLN is not fully specified: it is unclear whether the conditioning for scale and bias uses shared or separate MLPs.
- The CE evaluation relies solely on CheXbert mapping to 14 labels; there is no evaluation with complementary clinically oriented metrics (e. g. , RadGraph F1) or a human reader study, which limits external validity.
- The paper does not specify how CheXbert uncertainties are mapped (to negative/positive/ignored), nor whether CE scores are micro- or macro-averaged; these choices can materially alter P/R/F1.
- Validation on only one dataset may not be sufficient. How does the method generalize to other datasets (e. g. , CheXpert Plus, IU‑Xray) or to degraded image quality regimes? Any preliminary results?
- The paper lacks positioning with recent reliability-focused RRG works (e. g. , CURE [A]), making it unclear how it compares to alternative approaches for improving clinical correctness.
[A] Messina, Pablo, et al. “CURE: Curriculum-guided Multi-task Training for Reliable Anatomy Grounded Report Generation. “ arXiv preprint arXiv:2601.15408 (2026).
- Please rate the clarity and organization of this paper
Satisfactory
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission has provided an anonymized link to the source code, dataset, or any other dependencies.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(3) Weak Reject — marginally below the acceptance threshold, but would not mind if accepted, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The paper tackles an important problem and proposes a coherent framework; however, the main claim of clinical controllability is not convincingly validated. In addition, there are clarity and reproducibility issues (e.g., undefined clinical loss, missing CE details), and the evaluation is limited to a single dataset. Overall, I view the central idea as meaningful and the empirical signal as encouraging. With more clarifications (see weakness section) in the rebuttal, I’m willing to update my score.
- Reviewer confidence
Somewhat confident (2)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Accept
- [Post rebuttal] Please justify your final decision from above.
The rebuttal addressed most of my concerns, particularly regarding implementation details, CE evaluation protocol, and additional validation with RadGraph and CheXpert. These additions improve the paper’s clarity and empirical support.
However, I still find the terminology around the “clinical loss” somewhat unclear. The rebuttal clarifies that the method uses a CE-based reward within the RL objective rather than a separately defined clinical loss term, but this distinction is still not clearly reflected in the paper’s presentation.
Overall, the rebuttal resolved most of my concerns, although some clarity issues remain. I’m updating my score accordingly, with the hope that the authors will further clarify the formulation and terminology around the “clinical loss” in the final version.
Review #2
- Please describe the contribution of the paper
This paper proposes a reinforcement learning framework for precisio recall controllable RRG for chest X-rays, in which a continuous control parameter explicitly adjusts the balance between clinical precision and recall at inference time. This design allows the same model to flexibly generate reports that align with different clinical requirements, without retraining or changing decoding strategies. They further introduce a clinical loss that explicitly supervises the presence and absence of clinically relevant findings, enhancing clinical efficacy beyond language-only optimization. They adopt a group-relative training strategy that normalizes rewards within each sample group, effectively reducing variance and stabilizing policy optimization Experimental results show that their approach consistently outperforms state-of-the-art methods in both language quality and clinically oriented evaluation metrics, while providing reliable and interpretable control over the precision recall trade-off.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
This paper proposes a novel approach to enhance clinical consistency and controllability in RRG. The proposed method enables flexible precision recall control through a hybrid natural language and clinical efficacy reward mechanism, allowing the model to better capture clinically relevant findings while maintaining semantic coherence. Extensive experiments on the MIMIC-CXR dataset demonstrated the effectiveness and robustness of the approach compared with existing methods.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
The proposed method use GRPO which is used in DeepSeek. Recently, Nvidia released GDPO:Group reward-Decoupled Normalization Policy Optimization for Multi-reward RL Optimization. You may need to think about using this algorithm.
- Please rate the clarity and organization of this paper
Satisfactory
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission does not mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure reproducibility.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
This paper proposes a novel approach to enhance clinical consistency and controllability in RRG. The proposed method enables flexible precision recall control through a hybrid natural language and clinical efficacy reward mechanism, allowing the model to better capture clinically relevant findings while maintaining semantic coherence. Extensive experiments on the MIMIC-CXR dataset demonstrated the effectiveness and robustness of the approach compared with existing methods. Even though they do not use GDPO rather than GRPO, I think that they may expand their idea to choose GDPO easily.
- Reviewer confidence
Very confident (4)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
N/A
- [Post rebuttal] Please justify your final decision from above.
N/A
Review #3
- Please describe the contribution of the paper
1.An RL framework for radiology report generation that introduces a continuous control parameter \lambda to adjust the trade-off between clinical precision and recall at inference time, without need for retraining. 2.A hybrid reward function combining NLG metrics and clinical efficacy rewards (precision/recall weighted by \lambda), for joint optimization of linguistic quality and clinical correctness. 3.Evaluation on MIMIC-CXR showing SOTA performance on both NLG and clinical metrics, with smooth precision-recall controllability across \lambda values
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
1.The idea of controlling the precision-recall trade-off at inference time addresses a very important clinical need. Enabling this with a single model and a single parameter is a practical design. 2.The method achieves the best F1 and Recall among all methods in Table 1, with a large margin over the next best. The NLG metrics are also best across the board. The smooth, monotonic precision-recall curves in Fig. 2 demonstrate that \lambda provides meaningful and continuous control.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
1.All experiments are conducted exclusively on MIMIC-CXR dataset. Without evaluation on at least one additional dataset (e.g., CheXpert, or other dataset), it is not possible to assess whether the precision-recall controllability and performance gains transfer across cohorts or not. 2.All results in Tables 1 and 2 are single point estimates with no confidence intervals, standard deviations, or significance tests. Given that RL-based training is known to be sensitive to random seeds and sampling, this omission makes it difficult to determine whether the reported improvements over baselines are robust or not. Multiple runs with different seeds should be reported. 3.Table 2 ablates clinical loss and group-relative training, but does not ablate the PRC-AdaLN module. It is unclear how much of the controllability comes from the representation-level conditioning versus the reward-level \lambda weighting.
- Please rate the clarity and organization of this paper
Satisfactory
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission has provided an anonymized link to the source code, dataset, or any other dependencies.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(3) Weak Reject — marginally below the acceptance threshold, but would not mind if accepted, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The core idea of inference-time precision-recall control is well-motivated and the paper makes a novel methodological contribution. However, the single-dataset evaluation is a significant drawback. Without evaluation on at least one additional dataset with different characteristics, or imaging protocols, there is no evidence that the method generalizes. This is worsened by the lack of statistical testing across runs.
- Reviewer confidence
Very confident (4)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Accept
- [Post rebuttal] Please justify your final decision from above.
All the concerns that I had raised have been satisfactorily addressed by the authors, and therefore I revise my rating to “Accept”.
Author Feedback
We sincerely thank all reviewers for their constructive and encouraging feedback. We appreciate the recognition of our novel methodological contributions (R1, R2, R3, and Meta R2), the strength of our empirical results (R1, R2, R3, and Meta R2), and the clarity of our presentation (R1, R2, R3, and Meta R2). All clarifications, additional implementation details, and relevant discussions will be incorporated into the final version. We address each reviewer’s concerns below:
Method Details (Meta R2.Q3, R1.Q1, R1.Q2 and R1.Q4): The overall training loss consists of standard per-word cross-entropy loss and RL loss (eq. 3). The RL loss consists of natural language generation (NLG) reward and clinical efficacy (CE) reward (eq. 5). NLG reward combines BLEU-4, METEOR and ROUGE-L (eq. 6), and CE reward is based on precision/recall from CheXbert (eq. 2). In the PRC-AdaLN module (eq. 1, Fig. 1), the scale and biases were generated using separate MLPs. In both training and evaluation, uncertain CheXbert labels were mapped to positive, while negative and blank labels were mapped to negative. CE precision, recall, and F1 were micro-averaged over all 14 labels and test samples.
Additional Dataset (Meta R2.Q1, R1.Q5, R3.Q1): We evaluated our method on CheXpert Plus and it achieved better NLP and clinical metrics compared to R2GenCSR [1], which reported the SOTA performance on CheXpert Plus, and MedGemma. We report representative metrics including ROUGE-L as the selected NLG metric, CheXbert F1, and RadGraph F1.Our method achieved 0.225/0.502/0.169, outperforming R2GenCSR (0.220/0.475/0.159) and MedGemma (0.153/0.434/0.144).
Confidence Intervals (Meta R2.Q2 and R3.Q2): Confidence intervals and statistical tests were performed using paired bootstrapping with 10,000 repeats on existing MIMIC-CXR results and CheXpert Plus results. The 95% CIs for our method on MIMIC-CXR were [0.331, 0.336] for ROUGE-L, [0.564, 0.579] for CheXbert F1, and [0.237, 0.245] for RadGraph F1.The 95% CIs on CheXpert Plus were [0.224, 0.227], [0.497, 0.507], and [0.167, 0.171] for the three metrics, respectively. Paired bootstrap tests showed statistically significant improvement compared to MLRG, R2GenCSR, and MedGemma on corresponding datasets (p < 0.0001).
Additional Metrics (Meta R2.Q4 and R1.Q3): We calculated RadGraph F1 for MIMIC-CXR and our method achieved the best performance (0.240) compared to MLRG (0.227), and MedGemma (0.190), showing consistent clinically relevant improvement with newer metrics.
Ablating PRC-AdaLN module (Meta R2.Q5 and R3.Q3): The controllability comes from both the representation-level λ conditioning and reward level λ weighting. Without the representation-level conditioning, there is no way to control the precision/recall during inference without retraining; without the reward level weighting, the representation-level conditioning does not map to any meaningful outcomes. The testing-time controllability was only possible because the same λ was fed to the PRC-AdaLN module and used for reward during training. We also tried injecting λ differently, such as encoding λ and concatenating to image features. However, this strategy provided less controllability than our current design, since varying λ during inference produced no change in the generated reports.
Concurrent Works: (Meta R2.Q6, R1.Q6 and R2.Q1): Thanks for the suggestion and we will discuss recent concurrent works CURE [2] and GDPO [3] in the final version. CURE focuses on curriculum-guided anatomy grounding, while our method focuses on inference-time precision-recall controllability via λ-conditioned representation and λ-controlled clinical reward. We will explore integrating GDPO into our framework in future work.
Refs: [1] Wang et al., R2GenCSR, IEEE JBHI, 2026.[2] Messina, Pablo, et al. CURE, arXiv preprint arXiv:2601.15408 (2026). [3] Liu et al., GDPO, arXiv:2601.05242, 2026.
Meta-Review
Meta-review #1
- Your recommendation
Invite for Rebuttal
- Please justify your decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. In case of an invitation for rebuttal, clarify which points are important to address in the rebuttal.
The paper proposes a reinforcement learning framework for chest X-ray radiology report generation that enables explicit, inference-time control of the clinical precision–recall trade-off via a control parameter during inference. Experiments on MIMIC-CXR report improvements over prior methods in standard NLG and CE metrics, supported by ablations.
The problem setting is clinically meaningful and inference-time precision–recall controllability is a practical, well-motivated contribution. The paper is reasonably clearly written, and method shows strong empirical performance on MIMIC-CXR. However, several concerns need to be addressed to increase confidence in the claims, including: 1) Evaluation is restricted to a single dataset (MIMIC-CXR), and authors could provide results on at least one additional dataset or compelling evidence of robustness to distribution shifts. 2) Results are reported as single point estimates without confidence intervals, standard deviations, or significance tests. 3) There are missing or unclear methodological details that affect reproducibility and interpretation (e.g., definition of “clinical loss”, PRC-AdaLN implementation, CheXbert uncertainty mapping, metrics averaging). 4) Clinical validation is narrow (no complementary clinical metrics such as RadGraph, and no human evaluation). 5) Missing ablation of PRC-AdaLN leading to unclear impact of PRC-AdaLN vs. reward-level λ weighting. 6) Positioning of paper not clear relative to recent works (e.g. CURE, GDPO).
Overall, the core idea (single-model, inference-time control of the clinical precision–recall trade-off) is novel, relevant, and promising, but the current evidence does not yet meet the bar for acceptance. I invite the paper for rebuttal to address the reviewers’ concerns.
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation.
The rebuttal addressed several of the reviewers’ points including the implementation details, confidence intervals, positioning with prior work, and additional evaluation metrics. These clarifications improved reviewers’ confidence, leading two reviewers to update their recommendations to Accept. That said, I also share the AC’s concern that new results on additional dataset introduced in rebuttal should not be used as the basis for the final recommendation, and my assessment therefore does not rely on those additional experiments. Overall, the paper presents a meaningful contribution, and the rebuttal sufficiently resolved most of the key concerns raised on the submitted version.
Meta-review #2
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation.
The authors have addressed the reviewers’ comments well. All reviewers acknowledged the significance of the work. Therefore acceptance of the paper is concluded.
Meta-review #3
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Reject
- Please justify your recommendation.
The paper proposes a reinforcement-learning framework for radiology report generation that aims to enable controllable precision–recall trade-off behavior through a conditioning parameter used in both decoder modulation and clinical reward weighting. The work puts forward a clinically meaningful problem and reports strong performance with interpretable precision–recall trends on MIMIC-CXR. While the method demonstrates predictable precision–recall trade-off behavior through λ, the submitted evidence supports steerability of the operating point rather than reliable control of desired precision or recall targets. In addition, the reinforcement-learning objective explicitly optimizes a CheXbert-defined precision–recall target, while the main clinical evaluation also relies on CheXbert-based metrics, making complementary clinical validation beyond the CheXbert label space particularly important. Although rebuttal-only RadGraph results partially alleviate this concern, they are not part of the submitted study and are not considered in the decision. Finally, the submitted evidence is limited to a single dataset, making generalization beyond MIMIC-CXR less clear. Overall, the work presents a clinically meaningful and promising direction. Incorporating and improving the additional validation provided in the rebuttal and clarifying the above issues would substantially strengthen the work and make the study more convincing.