List of Papers Browse by Subject Areas Author List
Abstract
Automatic radiology report generation has the potential to assist radiologists by providing preliminary clinical descriptions. Despite recent progress in medical vision-language models, generated reports often contain factual inconsistencies and repetitive templated findings, indicating weak grounding to image evidence. Retrieval-augmented generation (RAG) improves grounding by conditioning on prior reports; however, existing end-to-end optimization suffers from representation inconsistency, as actively updated query representations diverge from stale historical keys. This mismatch destabilizes the retrieval space and disrupts supervision. We propose CARE (Clinically Aligned Retrieval Evidence), a multimodal framework introducing temporally coherent retrieval supervision. CARE maintains a shared memory queue populated by a slowly evolving momentum key encoder to ensure a stable representation space. We further introduce a disease-aware alignment objective that enforces diagnostic agreement between input images and retrieved reports to guide generation. Trained on IU-Xray and evaluated on the unseen MIMIC-CXR-RRG dataset, CARE significantly improves report quality. It nearly doubles clinical entity extraction performance in training-free adaptation settings and substantially reduces repetitive report collapse across multiple backbones. The source code is available at: https://github.com/Levii-zhu/CARE
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/3601_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/Levii-zhu/CARE
Link to the Dataset(s)
IU X-Ray dataset: https://drive.google.com/file/d/1c0BXEuDy8Cmm2jfN0YYGkQxFZd2ZIoLg/view
MIMIC-CXR-RRG dataset: https://huggingface.co/datasets/Yamini-1628/MIMIC-CXR-RRG
BibTex
@InProceedings{ZhuJin_CARE_MICCAI2026,
author = { Zhu, Jintao AND Gu, Yi AND Ma, Jiaxin},
title = { { CARE: Clinically Aligned Retrieval Evidence for Consistent Radiology Report Generation } },
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 authors propose CARE (Clinically Aligned Retrieval Evidence), a multimodal Retrieval-Augmented Generation (RAG) framework designed to improve the factual consistency of radiology reports. To solve the representation inconsistency caused by actively updated queries diverging from stale historical keys during end-to-end training, CARE implements a slowly evolving momentum key encoder coupled with a shared memory queue. Additionally, it introduces a disease-aware alignment objective to ensure that the retrieved evidence is diagnostically compatible with the input visual features.
- 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) Elegant Theoretical Motivation: The paper identifies a fundamental bottleneck in end-to-end RAG optimization—the destabilization of the retrieval space. Adapting the momentum update mechanism (inspired by contrastive learning paradigms like MoCo) to manage cross-modal retrieval queues is a technically sound and elegant solution. 2) Exceptional Zero-Shot Domain Adaptation: The framework exhibits remarkable out-of-domain generalization. On the unseen MIMIC-CXR benchmark, simply populating the external memory queue with target-domain text (without any image or label fine-tuning) nearly doubled the clinical entity extraction performance (RG-F1), highlighting its flexibility for real-world deployment. 3) Effective Mitigation of Mode Collapse: The N-gram diversity analysis effectively proves that CARE successfully breaks the “shortcut learning” loop. By anchoring generation to stabilized retrieved evidence, the model avoids regurgitating rigid, repetitive templates (e. g. , defaulting to “clear lungs”), resulting in richer and more clinically detailed reports.
- 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) Frozen LLM Limits Deep Adaptation: The framework freezes the core LLM parameters during training to save compute, optimizing only the query encoder and projection layers. While efficient, this approach may bottleneck the model’s ability to deeply internalize complex, institution-specific reasoning logic or highly specialized syntactic styles for longer reports. 2) Unverified Robustness to Noisy Retrievals: The system assumes the retrieval queue will return highly relevant and accurate evidence. However, real-world Electronic Health Records (EHR) are notoriously noisy. The paper lacks an ablation study or stress test demonstrating the generator’s robustness and self-correction capability when forced to condition on erroneous or contradictory retrieved reports. 3) Absence of Human Expert Evaluation: Although the automated clinical metrics (e. g. , RadGraph F1) show significant improvement, these metrics cannot fully capture clinical fluency, coherence, and clinical utility. The lack of qualitative assessment by expert radiologists slightly weakens the clinical impact claims.
- Please rate the clarity and organization of this paper
Good
- 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 authors claimed to release the source code and/or dataset upon acceptance of the submission.
- 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?
CARE provides a highly effective, theoretically grounded solution to a known problem in RAG-based report generation. The stabilization of the retrieval space via momentum encoders yields undeniable empirical benefits, particularly in zero-shot cross-dataset generalization. My score is slightly conservative (Weak Accept) because of the reliance on the rigid CheXbert labeler for alignment and the lack of an anonymized code repository for review. I look forward to the authors’ rebuttal regarding the ablation study on the unseen dataset and their plans for reproducibility.
- Reviewer confidence
Confident but not absolutely certain (3)
- [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 authors have addressed the majority of my concerns.
Review #2
- Please describe the contribution of the paper
Addresses a Real and Underexplored Problem (RAG Instability) and Identifies a subtle but important issue: representation inconsistency in end-to-end RAG training. Proposes a momentum memory queue to stabilise the retrieval space. This is a meaningful systems-level contribution, not just performance tuning.
- 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.
Well-Motivated Retrieval Design (MoCo-style for Multimodal RAG)
- Adapts momentum encoder + memory queue (from contrastive learning) to both multimodal retrieval and report generation. Clinically Meaningful Alignment Objective
- Introduces disease-aware alignment loss and enforces diagnostic consistency between input image and retrieved reports. Improves Generalization Demonstrates training-free adaptation to unseen dataset (MIMIC-CXR) with nearly 2× improvement in entity extraction (RG-F1).
- 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.
- Method is well-engineered but lacks strong conceptual novelty. Core components are known: MoCo-style momentum encoder RAG, contrastive learning, label alignment.
- Still Relies on Retrieval Quality (Upper Bound Issue), no mechanism to verify or filter incorrect retrieved evidence.
- No Fine-Grained Alignment: Alignment happens at report level (retrieval) not at entity-level grounding or region-level correspondence.
- Please rate the clarity and organization of this paper
Good
- 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?
Please see strength and weakness above. Overall, it’s an useful application in the clinical domain and improvements are clearly illustrated throughout the manuscript.
- 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.
concerns well addressed. recommend for acceptance.
Review #3
- Please describe the contribution of the paper
This paper presents a method for improving radiology report quality via a framework that alters the RAG of a VLM. A momentum encoder and a shared memory queue are added to the model, which smooths the retrieval representation space during optimization. Additionally, a disease-aware alignment objective is employed to encourage retrieval of cases with similar pathologies. This is implemented as a softmax relaxation in the label space and is paired with the query encoder to couple the retrieval with visual representation. The retrieved evidence is combined with image features and fed into an LLM to produce the output radiology report. The method is fine-tuned on the IU-Xray dataset and evaluated on both the IU-Xray dataset and the unseen MIMIC-CXR-RRG dataset. Nine metrics are employed for comparison: Mi-F1, Ma-F1, RG-F1, RG_ER, and CliQ0 for the clinical metrics, and B-1, B-4, MTR, and R-L for the lexical 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 generalization study of training on IU-Xray and assessing on MIMIC-CXR-RRG is appreciated to assess the ubiquity of the method
Multiple models are employed for testing the method (LLaVA-Med and LLaVA-Rad), which further helps to give generalized performance measures, as LLMs are known to produce wildly varying results depending on the system prompt and employed model.
The provided ablation study justifies the inclusion of the various proposed components.
- 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.
Of all the employed metrics, the ones that are most relevant to clinical utility appear to be RG-F1 and CliQ0.However, the model consistently underperforms in these categories compared to other methods. Thus, the claims by the authors for the utility of the method are dubious. It is unclear to this reviewer how any of the other proposed benefits of the model, such as the N-grams in Figure 2 or the lexigraphical metrics, can outweigh the impact of superior clinical accuracy metrics.
An opportunity is missed to compare the proposed method fine-tuned on the MIMIC-CXR dataset to compare with state-of-the-art performance models. The provided method demonstrates highly substandard performance to the fine-tuned methods. While this is somewhat expected, it could have been counterbalanced with a fine-tuned example.
The CheXbert F1 score is mentioned on page 5, but does not appear to be reported in any of the tables.
- 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 authors claimed to release the source code and/or dataset upon acceptance of the submission.
- 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
The labels in figure 2 are too small and should be increased for readability.
The image in Figure 3 is formatted poorly and should be larger for visibility.
The paper uses the term “hallucinations” twice. The term is highly ambiguous can refer to several different items; see https://pubmed.ncbi.nlm.nih.gov/41811081/ for a overview of some of the definitions (although the work unfortunately lacks citations for many of the employed definitions). Various definitions have been posed by entities such as OpenAI (https://storage.prod.researchhub.com/uploads/papers/2020/06/01/language-models.pdf) and (very recently) the FDA (https://doi.org/10.1016/j.ailsci.2025.100145). To promote scientific clarity and preciseness, I suggest that the term hallucination be changed to a more precise phrasing for the two occurrences in the paper, or an explicit definition of the term with a relevant citation provided in the introduction.
- 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?
My impression of this paper is that it fails to improve performance compared to other approaches on the most relevant metric for clinical tasks. Compared to RG-F1 and CliQ0, every other metric appears to be a distraction. Furthermore, overall performance is substantially substandard compared to fine-tuned models, which further restricts the applicability of the method.
- Reviewer confidence
Somewhat confident (2)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Reject
- [Post rebuttal] Please justify your final decision from above.
I thank the authors for their response to the provided feedback, particularly the clarifications about the CheXbert F1-score. However, this new information only leaves me with a sense that the results are muddled and inconclusive.
After reading the other reviewer’s comments, the impression I am left with of this paper at the end of the review process is that it is an incremental advancement with limited novelty aimed at a functional application. However, the desired function (training-free cross-domain adaptation in the medical domain) retains astoundingly low performance. It is unclear if the proposed method will be of any utility when such a solution becomes practical, as it is likely that the underlying technology must change a substantial amount before reaching usability.
Since the authors mention that they are targeting practical clinical deployment in their work from the second line in their rebuttal, I lean towards rejection as I believe that they have failed to make meaningful steps towards that goal. Sufficient novelty would persuade me that progress is being toward a more permanent solution, but the current contribution, to me, appears transient in nature.
Author Feedback
We thank the reviewers for their highly constructive feedback and recognition of CARE’s theoretical motivation. Our work resolves systems-level retrieval instability to enable robust, training-free cross-domain adaptation for practical clinical deployment. Primary concerns are addressed below.
Q1: (R1, R2) Robustness to noisy retrieved evidence. Although CheXbert pseudo-labels are standard in prior work, CARE mitigates their rigidity through soft retrieval-guided conditioning rather than hard retrieval supervision. Retrieved candidates receive adaptive soft weights (α) based on continuous visual-textual similarity, naturally suppressing noisy or incompatible evidence during alignment. Generation is jointly conditioned on visual features and weighted retrieval evidence, while supervision still comes from the ground-truth report, reducing over-reliance on noisy retrievals. We will note the lack of adversarial stress-testing as a limitation.
Q2: (R3) Concerns regarding slightly lower RG-F1 and CliQ0 performance. We agree that RG-F1 and CliQ0 are important clinically grounded metrics. However, these global overlap metrics may not fully capture sensitivity to less frequent pathological findings. Because normal cases dominate the training distribution, standard SFT models often converge toward repetitive “normal” templates, maintaining strong report-level scores while underdetecting abnormalities. This behavior is evident in LLaVA-Rad (SFT), which achieves high RG-F1 (38.0) but lower pathology-sensitive performance (5Mi/5Ma-F1: 34.4/23.2), and is even more severe in LLaVA-Med (1.0/0.3). In contrast, CARE stabilizes retrieval grounding and consistently improves abnormality sensitivity across both backbones (35.1/26.9 and 20.4/15.4). Our diversity analysis is intended as complementary evidence that CARE mitigates repetitive template collapse rather than memorizing dominant normal patterns. We will clarify this distinction more explicitly in the revised paper.
Q3: (R3) Missing comparison to MIMIC-CXR-fine-tuned version. We would like to clarify that while supervised models naturally achieve higher MIMIC-CXR scores, CARE targets a fundamentally different setting: training-free cross-domain adaptation. In clinical environments, institutions often lack the computational resources or annotated data required for repeated retraining. Under this zero-shot constraint, adapting the retrieval memory via CARE demonstrates significant effectiveness, improving RG-F1 from 7.0 to 13.0 without any target-domain training. We will clarify this in the revised paper.
Q4: (R2) Conceptual novelty. While the core components (e.g., MoCo) are established, our novelty is functional rather than strictly architectural: resolving an underexplored systems-level instability in multimodal RAG. Beyond expanding the negative pool, CARE leverages the momentum queue to maintain a temporally stable retrieval space. Prior systems assume a static space; updated queries drift from stale queue keys, destabilizing LLM retrieval supervision. CARE stabilizes this process, ensuring reliable supervision under distribution shifts. We will clarify this stabilization perspective more explicitly in the revised manuscript.
Q5: (R1) Frozen LLM. We froze the LLM to isolate our retrieval stabilization module and maintain a minimal training footprint. This bottlenecks internalizing institution-specific syntactic styles; integrating PEFT (e.g., LoRA) will be outlined as a next step.
Q6: (R3) Missing definition of “CheXbert F1 score”. We apologize for the missing definition of “CheXbert F1 score”, which we refer as 5Mi-F1 and 5Ma-F1 metrics in tables 1 and 2.We will clarify this in the dataset and evaluation metrics section.
Other comments: We will improve figures qualities with more enlarged labels and clearer format (R3). We will also rephrase “hallucination” to “factual inconsistency” for a more precise expression (R3). Our code will be released upon acceptance (R1, R2).
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 addresses the instability of retrieval-augmented generation in radiology. Existing RAG frameworks often suffer from “representation drift,” where updated query encoders diverge from static historical keys in the retrieval database. CARE introduces a momentum-based retrieval supervision mechanism to maintain a stable representation space. This is paired with a disease-aware alignment objective to ensure that retrieved reports provide a consistent diagnostic prior for the language model decoder.
While the paper presents good contributions, its recommendation for acceptance is contingent upon the authors clarifying the following points during the rebuttal phase:
1.A significant concern is the model’s reliance on the quality of retrieved evidence. Reviewers noted that the framework lacks a mechanism to filter or “self-correct” when the retrieval queue returns erroneous, noisy, or contradictory reports from the EHR.
2.While the paper highlights improvements in linguistic metrics, reviewers pointed out that the model underperforms in categories most relevant to clinical utility, such as RadGraph F1 and RadCliQ. The reviewer argued that lexical diversity or N-gram improvements cannot outweigh the impact of superior clinical accuracy, making the authors’ claims regarding clinical utility questionable.
3.The framework is seen as well-engineered but conceptually incremental. It relies on established components like MoCo-style momentum encoders and standard contrastive learning. Furthermore, alignment occurs only at the report level. The model lacks entity-level grounding or region-to-text correspondence, which limits its ability to pinpoint specific visual findings.
- 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. Two reviewers acknowledged significance of the work. Therefore acceptance of the paper is concluded.
Meta-review #2
- 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 scores are still mixed after the rebuttal. Although CARE is well motivated, I recommend rejection. The contribution is incremental, adapting known momentum-queue and contrastive ideas to RAG, while lacking entity- or region-level grounding and robust filtering of noisy retrieved evidence. Most importantly, the method underperforms on key clinical utility metrics such as RG-F1 and CliQ0, and the rebuttal does not convincingly establish practical clinical value or meaningful progress toward deployment.
