Paper Info Reviews Author Feedback Meta-Review Back to top
List of Papers Browse by Subject Areas Author List

Abstract

Automatic medical report generation (MRG) has significant clinical value, as it can alleviate the heavy workload of report writing for radiologists. Recent studies adopt retrieval-based strategies to inject external diagnostic knowledge into report generation. However, reliable MRG remains challenging because existing methods often retrieve diagnostically inconsistent reports and lack mechanisms for selective visual-text integration. In this work, we propose DCAF, a Disease-Consistent Visual-Text Alignment and Fusion framework for reliable radiology report generation. DCAF incorporates a Disease-Consistent Retrieval Enhancer that integrates disease priors via multi-positive contrastive learning and a disease-matching constraint to improve diagnostic consistency during retrieval. To better mimic the selective diagnostic reasoning process of radiologists, we further design an Optimal Transport-Enhanced Cross-Modal Fusion module that selectively integrates clinically relevant textual knowledge. In addition, an auxiliary disease-aware feature extraction mechanism encodes predicted diagnostic hypotheses into visual representations, enabling hypothesis-driven visual reasoning. Extensive experiments on IU X-Ray and MIMIC-CXR demonstrate state-of-the-art performance in both language generation quality and clinical efficacy.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/4526_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/shaojian0514/DCAF-RRG

Link to the Dataset(s)

IU X-Ray Dataset: https://openi.nlm.nih.gov/ MIMIC-CXR Dataset: https://physionet.org/content/mimic-cxr/

BibTex

@InProceedings{ShaJia_DiseaseConsistent_MICCAI2026,
        author = { Shao, Jianjun AND Tian, Yun},
        title = { { Disease-Consistent Visual-Text Alignment and Fusion for Reliable Radiology Report Generation } },
        booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
        year = {2026},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 16879},
        month = {September},
        page = {pending}
}

Reviews

Review #1

  • Please describe the contribution of the paper

    The paper proposes DCAF, a retrieval-augmented radiology report generation framework designed to improve diagnostic reliability. The method combines (i) a Disease-Consistent Retrieval Enhancer with disease-prior multi-positive contrastive learning and a disease-matching constrained retrieval objective, (ii) an Optimal Transport-Enhanced Cross-Modal Fusion module for selective use of retrieved textual knowledge, and (iii) a Disease Classifier for disease-aware feature extraction. On MIMIC-CXR and IU X-Ray, the method reports improved natural language generation metrics and CheXbert-based clinical efficacy metrics relative to the compared baselines.

  • 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 problem setting is clinically meaningful and well motivated. The manuscript identifies a concrete limitation of retrieval-augmented medical report generation: semantically similar reference reports may still be diagnostically inconsistent, which can reduce the reliability of the generated report. 2.The proposed framework is coherent and aligned with the stated objective. The Disease-Consistent Retrieval Enhancer explicitly incorporates disease priors into image-text alignment and retrieval; the Optimal Transport-Enhanced Cross-Modal Fusion module is intended to selectively use clinically relevant retrieved text rather than treating all retrieved tokens equally; and the auxiliary Disease Classifier provides an explicit disease-aware learning signal for the visual representation. 3.The empirical results are consistently strong on the reported benchmarks. On MIMIC-CXR, the method improves BLEU-4 from 0.137 (STREAM) / 0.136 (KIA) to 0.146 and clinical efficacy F1 from 0.499 (EKAGen) to 0.552.On IU X-Ray, BLEU-4 improves from 0.098 (PromptMRG) to 0.105 and F1 from 0.211 to 0.256.These gains are meaningful for this task and are observed on both language and clinical proxy metrics. 4.The experimental section includes useful supporting analyses. The ablation study on MIMIC-CXR shows incremental gains when DMCL, DCR, OTCF, and DFE are added, the top-k study provides some sensitivity analysis for retrieval, and the qualitative example is consistent with the paper’s claim that the proposed framework reduces clinically inconsistent statements.

  • 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.The methodological novelty appears moderate rather than substantial. DCL [12] already uses contrastive learning for chest X-ray report generation; PromptMRG [10], EKAGen [2], STREAM [25], and KIA [27] already incorporate diagnosis, knowledge, or retrieval signals into report generation; and the OT-based attention mechanism is adapted from MPSA / OT-Seg [26]. The current contribution is therefore best characterized as a task-specific integration of several reasonable ideas, and the paper would benefit from a clearer positioning of what is fundamentally new beyond the combination. 2.The central claim of reliability is supported mainly by automatic proxy metrics. Clinical efficacy is evaluated only through CheXbert-based precision, recall, and F1 on 14 labels. No radiologist study, factual error taxonomy, or patient-level safety analysis is provided. For a paper emphasizing ‘reliable’ report generation, stronger clinical validation would materially strengthen the claim. 3.The paper uses CheXbert-derived disease labels both to supervise disease-consistent retrieval and to evaluate the generated reports. Because both the retrieval supervision and the downstream clinical efficacy metric depend on the same report-derived automatic labeler, there is a risk that proxy alignment and label noise partially drive the reported gains. This limitation should be discussed more explicitly. 4.Reproducibility is only partial in the current version. The manuscript does not provide an anonymized code link, and several implementation details are insufficiently specified, including the loss weights (lambda_match, lambda_cls, lambda_gen), retrieval queue size and update policy, the exact top-k setting used on each dataset during training and inference, the handling of uncertain labels from CheXbert, and the precise integration point of MPSA within the generation architecture. 5.Some experimental and presentation choices need clarification. The IU X-Ray evaluation follows PromptMRG [10] by directly transferring a model trained on MIMIC-CXR to the entire IU X-Ray dataset, which is useful for transfer evaluation but is not equivalent to a standard in-domain benchmark. In addition, Eq. (2) and Eq. (4) would benefit from clearer dimensional and probabilistic explanations, and Table 1 contains minor presentation issues such as the ‘Precison’ typo.

  • 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?

    The paper addresses a relevant and timely problem in radiology report generation and presents a technically coherent framework. The empirical improvements on MIMIC-CXR and IU X-Ray are consistent across both language generation metrics and clinical proxy metrics, and the ablation study supports the usefulness of the main components. These are positive aspects. At the same time, the methodological novelty is moderate, the reliability claim is supported mainly by automatic report-derived proxies, and the reproducibility section should be strengthened with additional implementation details. Overall, I view the paper as marginally above the acceptance threshold for MICCAI, with the final assessment benefiting from clarification during rebuttal.

  • Reviewer confidence

    Confident but not absolutely certain (3)

  • [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.

    The rebuttal resolves some clarification issues, especially regarding the IU X-Ray zero-shot protocol and reproducibility details. However, the core contribution still appears to be an integration of existing ideas. The reliability claim remains mainly supported by CheXbert-based automatic metrics, while no radiologist study, factual-error analysis, or independent clinical evaluation is provided. Since CheXbert-derived labels are also used in the method, the proxy-coupling concern remains. Therefore, I do not think the paper is above the acceptance threshold.



Review #2

  • Please describe the contribution of the paper

    This paper proposes a retrieval-augmented radiology report generation framework (DCAF) to address disease-level inconsistencies in retrieval and cross-modal fusion. The three core contributions are: (1) a Disease-Consistent Retrieval Enhancer (DRE) combining multi-positive contrastive learning—supervised by CheXbert-derived labels—with a disease-matching constraint to enforce diagnostic consistency; (2) an Optimal Transport-Enhanced Cross-Modal Fusion (OTCF) module using Multi-Prompt Sinkhorn Attention (MPSA) to selectively weight textual evidence; and (3) an auxiliary disease classifier that injects predicted disease probabilities back into the visual representation space. The authors claim that their framework achieves state-of-the-art performance on MIMIC-CXR and IU X-Ray datasets.

  • 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.Well-identified and clinically grounded problem. The paper targets three concrete failure modes of existing retrieval-augmented methods: inconsistent retrieval, uniform cross-modal fusion, and the absence of disease semantics. The modular design makes the contribution highly decomposable and ablation-friendly.

    2.Multi-positive contrastive learning with disease-label supervision. As the authors claimed, using CheXbert-extracted labels to define positive sets in the contrastive objective (Eq. 1) is a principled improvement over standard single-positive CLIP-style learning. This directly addresses retrieval consistency and provides the largest single-module gain (+0.034 F1) in the ablation.

    3.Optimal Transport for selective cross-modal fusion. Applying MPSA to enforce globally constrained attention weighting is a methodologically sound approach that better models radiologists’ selective integration of reference reports.

    4.Comprehensive ablation. Table 2 provides a clean incremental ablation over all components, with consistent monotonic gains, lending credibility to each module’s individual contribution.

  • 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.Reliance on CheXbert labels introduces systematic bias. The DRE module relies entirely on CheXbert pseudo-labels for training supervision. Furthermore, the CE metrics are evaluated using the same CheXbert tool. This introduces a severe circular dependency; the high CE scores may reflect overfitting to CheXbert’s inherent biases rather than genuine clinical correctness.

    2.IU X-Ray evaluation protocol is methodologically weak. The model is trained on MIMIC-CXR and evaluated directly on the entire IU X-Ray dataset (Section 3), including training samples. This is non-standard and potentially introduces data leakage. Standard practice (e.g., R2Gen, M2KT) is to train and test within IU X-Ray using its own splits. The paper should either follow standard IU X-Ray splits or clearly justify the cross-dataset zero-shot evaluation protocol, as the current setup makes results on IU X-Ray difficult to interpret.

    3.Limited Methodological Novelty: The core components are essentially direct applications of existing techniques. The multi-positive contrastive learning is a direct adaptation of SupCon (Khosla et al., NeurIPS 2020 [1]), which has already been explored in similar contexts (e.g., PromptMRG (Jin et al., AAAI 2024 [2]), DCL(Li et al., CVPR 2023 [3])). The MPSA module is explicitly imported from OtSeg (Kim et al., ECCV 2024 [4]). The disease-aware feature injection (DFE) is conceptually very similar to the knowledge-adaptive mechanism in M2KT(Yang et al., MedIA 2023 [5]).

    4.Absence of LLM-Based Baselines: The baseline pool consists largely of pre-2024 encoder-decoder models. Modern methods leveraging Large Language Models (e.g., LLaVA-Med, MedGemma) demonstrate substantially higher performance, making the absolute “SOTA” claim overstated.

    5.Computational Overhead of k=18: Table 3 shows the model peaks at $k=18$ retrieved reports. Retrieving 18 references and attending over them with iterative Sinkhorn algorithms is computationally expensive. No inference cost or latency analysis is provided.

    [1] Khosla, Prannay, et al. “Supervised contrastive learning.” Advances in neural information processing systems 33 (2020): 18661-18673. [2] Jin, H., Che, H., Lin, Y., Chen, H.: PromptMRG: Diagnosis-driven prompts for medical report generation. In: Proceedings of the AAAI Conference on Artificial Intelligence, pp. 2607–2615 (2024)

    [3] Li, M., Lin, B., Chen, Z., Lin, H., Liang, X., Chang, X.: Dynamic graph en- hanced contrastive learning for chest X-ray report generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 3334–3343 (2023)

    [4] Kim, K., Oh, Y., Ye, J. C.: OtSeg: Multi-prompt Sinkhorn attention for zero-shot semantic segmentation. In: European Conference on Computer Vision, pp. 200– 217.Springer, Cham (2024)

    [5] Yang, S., Wu, X., Ge, S., Zheng, Z., Zhou, S.K., Xiao, L.: Radiology report gen- eration with a learned knowledge base and multi-modal alignment. Medical Image Analysis 86, 102798 (2023)

  • 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 provide sufficient information for 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

    1.The IU X-Ray evaluation protocol must be clarified or corrected. If the entire IU X-Ray dataset is used as a test set for a model trained on MIMIC-CXR, this must be stated explicitly and compared fairly against other zero-shot baselines. Several baselines in Table 1 (e.g., R2Gen, M2KT) were evaluated under a different protocol.

    2.Please report all loss weighting coefficients (λ_match, λ_cls, λ_gen) and the temperature τ to enable reproducibility.

    3.Consider reporting GREEN score or RadGraph-F1 in addition to CheXbert-based CE metrics, as these are increasingly adopted as more clinically meaningful evaluation metrics in radiology NLP.

    4.A comparison with at least one LLM-based report generation baseline (e.g., MedGemma, LLaVA-Med) would significantly strengthen the SOTA claim.

    5.Please discuss the computational overhead of retrieving k=18 reports at inference time, particularly in comparison to baselines that use smaller k values.

    6.An analysis of CheXbert label noise and its downstream impact on DRE module performance would strengthen the clinical reliability argument.

  • 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?

    DCAF presents a well-ablated framework with genuine methodological contributions (multi-positive contrastive learning with disease priors, OT-based selective fusion). However, the submission suffers from critical methodological flaws: the IU X-Ray comparison mixes zero-shot with fully-supervised baselines, invalidating the results, and the exclusive reliance on CheXbert for both training and evaluation introduces a severe circular bias. Furthermore, the core algorithmic novelty is marginal, as the modules are direct adaptations of existing works (SupCon, OtSeg, M2KT) without adequate contextualization. These issues fundamentally undermine the experimental conclusions. The paper requires major revisions in evaluation protocol and baseline comparisons before it can be considered for acceptance.

  • 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.

    The authors provided a clear rebuttal that addressed my major concerns, justifying an update of my decision to acceptance. (1) The authors successfully clarified that the IU X-Ray baselines inherited the exact same zero-shot protocol from prior work, confirming that the cross-dataset comparison was indeed fair and apples-to-apples. (2) Furthermore, they correctly pointed out that STREAM (one of their baselines) is an LLM-based method (TinyLlama-1.1B), which addresses my concern regarding the absence of LLM-based comparisons. (3) Regarding reproducibility and complexity, all missing hyperparameters (loss weights, temperature, queue size) were provided, and the computational complexity of the Sinkhorn iterations was adequately explained to be on the same order as MHCA. (4) Finally, while the inherent circularity of using CheXbert as a proxy evaluator remains a field-wide limitation, the authors made a fair argument that their improvements are not merely overfitting to CheXbert biases, as evidenced by consistent gains in independent NLG metrics like BLEU-4.Overall, the rebuttal demonstrated that the framework is methodologically sound and the experimental comparisons are fair. The system-level novelty of integrating DMCL, MPSA, and DFE to address specific breaks in the ARRG pipeline is convincing. I now recommend acceptance.



Review #3

  • Please describe the contribution of the paper

    The paper proposes DCAF, a retrieval-augmented radiology report generation framework built around three components: (1) a Disease-Consistent Retrieval Enhancer (DRE) that combines multi-positive contrastive learning (DMCL), where reports sharing identical CheXbert-derived disease labels serve as multiple positives, with a disease-matching constraint (DCR) that minimizes cross-entropy between the disease annotations of retrieved references and the query image; (2) an Optimal Transport-Enhanced Cross-Modal Fusion (OTCF) module based on Multi-Prompt Sinkhorn Attention (MPSA), which replaces standard multi-head cross-attention to selectively weight retrieved textual features conditioned on visual evidence; and (3) a Disease Classifier (DFE) that applies focal-loss-supervised classification on fused features and re-injects predicted disease probabilities back into the visual representation. The complete system is trained with a composite loss combining contrastive, retrieval consistency, classification, and generation objectives. Experiments on MIMIC-CXR and IU X-Ray report improvements over a comprehensive set of baselines across both NLG and Clinical Efficacy 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.

    1.The paper identifies three specific, discrete failure modes in retrieval-based MRG — diagnostic inconsistency in retrieved references, coarse cross-modal fusion, and the absence of disease semantics in the visual representation — and addresses each with a targeted module. This one-to-one correspondence between problem and solution is methodologically clean and makes the paper’s contributions individually legible. 2.Standard single-positive contrastive objectives are poorly matched to radiology, where multiple reports may be semantically and clinically equivalent for the same disease profile. The use of CheXbert-derived multi-label annotations to define positive sets is a principled and domain-appropriate adaptation of contrastive learning. The distinction from prior work (e.g., PromptMRG, EKAGen) that uses CLIP-similarity-based retrieval without disease-level supervision is meaningful and clearly articulated. 3.Equation (2) uses average binary cross-entropy between input and retrieved disease annotations as a retrieval penalty — a lightweight, interpretable, and differentiable constraint that can be directly integrated into joint training. This avoids the engineering overhead of separate re-ranking pipelines and is clearly novel within the ARRG literature. 4.The paper compares against ten baselines covering a broad range of paradigms (memory-driven, knowledge-graph, classification-guided, retrieval-augmented), with performance on both MIMIC-CXR and cross-dataset IU X-Ray evaluation. The CE-F1 improvement of 0.552 over the next best (0.499 for EKAGen) on MIMIC-CXR is substantial and consistent with the framework’s clinical grounding objective. The cross-dataset evaluation, while following the established PromptMRG protocol, adds practical significance. 5.Table 2 incrementally adds each component in a fixed order and reports consistent monotonic improvements. The largest single gains are associated with DCR (CE-F1: 0.471 → 0.524) and DFE (CE-F1: 0.531 → 0.552), directly validating the paper’s core claims about disease consistency and disease-aware feature learning.

  • 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.The cross-modal fusion module is based directly on MPSA from OtSeg (Kim et al., ECCV 2024), a method designed for zero-shot semantic segmentation. The paper’s contribution here is the application of this existing module to cross-modal medical report generation rather than a novel formulation. The theoretical justification for why OT-based attention is specifically appropriate for this task — beyond the qualitative analogy to radiologists’ selective reasoning — is not formally developed. How does MPSA enforce global constraints in a way that standard MHCA cannot for this specific retrieval-fusion use case? This remains unanswered and limits confidence in the module’s contribution. 2.Table 3 reports that retrieval performance peaks at k=18 (BLEU-1: 0.417, F1: 0.256), which is substantially larger than the k values typically used in retrieval-augmented generation (most prior work uses k=3 to k=10). With k=18 retrieved reports contributing to the disease-matching constraint in Eq. (2), the computational overhead and the risk of averaging over highly heterogeneous disease profiles are non-trivial. The paper does not analyze the diversity of retrieved references at k=18, nor does it discuss the sensitivity of γ to noisy or conflicting disease annotations among the k retrieved samples. 3.The formulation in Section 2.1 introduces a “learnable disease-aware projection layer” yielding embeddings F_I and F_T in R^{d×e}. The notation implies that d disease-category-specific projections are applied, but the exact architecture and whether d independent projection heads or a single projection followed by reshaping is used is not described. This is a non-trivial architectural choice that affects the inductive bias of the DMCL objective and is critical for reproduction. 4.The paper evaluates models trained on MIMIC-CXR directly on the full IU X-Ray dataset (without a separate training split), following PromptMRG. While this is a recognized protocol, it introduces a confound: performance gains on IU X-Ray may reflect better MIMIC-CXR training rather than cross-dataset generalization per se. Furthermore, the CE metrics on IU X-Ray (F1: 0.256 for DCAF) are markedly lower than on MIMIC-CXR (F1: 0.552), and the paper does not analyze this gap or whether disease distributions in IU X-Ray are well-calibrated with the MIMIC-trained CheXbert labeler used for evaluation. 5.Equation (7) defines the training objective as a weighted combination of four losses (L^D_con, γ, L_cls, L_gen) with coefficients λ_match, λ_cls, and λ_gen. None of these values are reported in the Implementation Details section, nor is any sensitivity analysis provided. Given that the balance between retrieval consistency and generation losses is central to the paper’s claims, this omission is a reproducibility concern. 6.In Fig. 2, the generated report (Ours) contains the sequence “The cardiomediastinal silhouette is normal,” which is highlighted in red as contradicting the ground truth (which describes a mildly enlarged cardiac silhouette). This is a clinically significant hallucination that occurs despite the paper’s primary claim of improved clinical consistency. While a single example does not refute the aggregate results, the authors do not acknowledge or analyze this error, which weakens the credibility of the qualitative evaluation.

  • 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

    1.Could the authors clarify what specific property of MPSA, beyond the general OT framing, makes it more appropriate for retrieval-text cross-modal fusion than, for example, a gated attention mechanism or a standard cross-attention with learned relevance scores? Is there any diagnostic evidence (e.g., attention weight visualizations) showing that MPSA suppresses irrelevant retrieved content more effectively than MHCA?

    2.On the large k=18: What is the average number of distinct disease profiles represented in the k=18 retrieved reports? Does the disease-matching constraint γ remain well-behaved when retrieved reports contain conflicting annotations across their 14 disease dimensions?

    3.On the loss hyperparameters: What values were used for λ_match, λ_cls, and λ_gen, and were these determined via grid search or validation-set tuning?

    4.On the qualitative contradiction: The generated report in Fig. 2 states “The cardiomediastinal silhouette is normal” while the ground truth indicates a mildly enlarged cardiac silhouette. Could the authors comment on whether this type of contradiction persists across the test set and whether the CE metric captures this class of error?

    5.On the disease-aware projection layer: Could the authors clarify whether F_I ∈ R^{d×e} represents d independent disease-specific projections of the image embedding, or a single projection reshaped to d×e? The distinction affects the expressivity of the DMCL objective.

  • 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?

    DCAF presents a coherent and clinically motivated framework for retrieval-augmented report generation with well-targeted individual contributions. The multi-positive contrastive learning and disease-matching retrieval constraint are principled and novel components backed by strong ablation evidence, and the CE-F1 improvement on MIMIC-CXR is meaningful. However, the paper’s rating is held at Weak Accept rather than Accept due to several concerns that require clarification. The OTCF module borrows directly from an existing segmentation method without sufficient domain-specific justification, the loss hyperparameters are unreported, the disease-aware projection architecture is underspecified, and the large k=18 retrieval setting is unexplained. The uncorrected clinical contradiction in the qualitative figure, while not fatal, is an indicator that the clinical reliability narrative is overstated in places. These issues are collectively addressable in a rebuttal, and the paper’s overall contribution is likely above the rejection threshold if the authors can provide adequate clarification.

  • 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.

    N/A



Author Feedback

We thank the reviewers for the constructive feedback. Concerns cluster into these themes.

C1.IU X-Ray protocol [R1.W5, R2.W2+C1, R3.W4]. DCAF is trained only on MIMIC-CXR; IU is zero-shot (no gradient, no patient overlap, no leakage). We follow PromptMRG’s “MIMIC-train → full-IU test” protocol because the standard IU test split has very few positives for several diseases (e.g., pneumonia/consolidation only ~2 each), making disease-aware CE unstable [10, p.5, l.325–330]. We respectfully clarify a misunderstanding in R2.W2: our Tab.1 does not mix protocols; its IU baselines inherit the zero-shot protocol from PromptMRG[10] Tab.1’s † re-evaluation, not their original in-domain IU splits. Thus DCAF IU F1=0.256 is apples-to-apples. The MIMIC→IU gap reflects IU’s normal-class skew, domain shift, and off-MIMIC CheXbert calibration, not leakage/protocol mixing.

C2.Novelty beyond direct adaptation [R1.W1, R2.W3, R3.W1+C1]. DCAF’s novelty is problem-formulation level: a disease-consistency-driven retrieval–fusion–feature loop for ARRG. Prior retrieval MRG assumes semantic≈diagnostic consistency, so generators inherit wrong diagnoses from conflicting reports. Each stage fixes a distinct break: (a) DMCL positives=disease-vector equality, enforcing diagnostic—not merely semantic—retrieval; SupCon/DCL/PromptMRG cannot. (b) MPSA (adapted from OtSeg[26]) imposes doubly-stochastic token budgets — a global constraint MHCA/local/gated attention cannot enforce. (c) DFE feeds explicitly supervised focal ŷ back via F_d=ŷ·Φ+F_c, unlike M2KT latent memory. Tab.2 shows each block contributes (DMCL +0.034, DCR +0.053, DFE +0.021 CE-F1); novelty is the DMCL→MPSA→DFE loop, not any single module.

C3.CheXbert [R1.W2-W3, R2.W1+C6, R3.W6]. We agree CheXbert-based CE is a proxy and may introduce proxy-level coupling. However, DCAF does not optimize CE nor regress to CheXbert outputs: CheXbert only provides discrete grouping/consistency signals for DMCL/DCR, while generation is supervised by GT reports. If gains were mainly CheXbert-bias artifacts, we would expect the improvements to concentrate on CE; instead, CheXbert-independent BLEU-4 also improves (STREAM 0.137→DCAF 0.146, +6.6%), and the same evaluator scores all baselines. CheXbert is an ARRG-standard labeler (PromptMRG/DCL/EKAGen; macro-F1≥0.85 on MIMIC-CXR); DMCL dilutes noisy positives via multi-positive sets and DCR averages BCE over 14 labels. Fig.2: cardiomegaly is captured (gen: “enlarged”; GT: “mildly enlarged”), but “cardiomediastinal silhouette is normal” remains a sentence-level contradiction. DCAF reduces, not eliminates, fine-grained errors; GREEN/RadGraph-F1 and reader studies are future work.

C4.Reproducibility [R1.W4, R2.C2, R3.W3+W5+C3+C5]. Code/configs will be released upon acceptance. Hyperparams: λ_match=0.5, λ_cls=λ_gen=1.0 (held-out validation); τ=0.07; MoCo FIFO queue 65,536 (m=0.999); top-k=8/12/18 (train/MIMIC/IU); CheXbert-uncertain→negative (per PromptMRG); MPSA=3 Sinkhorn iters in decoder cross-attn. R3.W3: disease-aware projection = d independent disease-specific linear heads ∈R^e stacked into R^{d×e} (per-disease bias, NOT reshaped single projection).

C5.k=18 & inference [R2.W5+C5, R3.W2+C2]. k=18 is an upper bound (~6–8 distinct disease profiles per query under DMCL clustering), not fixed concatenation. DCR computes per-disease BCE over 14 labels, tolerant to partial conflicts; MPSA downweights irrelevant text. Table 3 plateaus at k=12/18/24 (F1≈0.25), so k=12 is a lower-cost option. With 3 Sinkhorn iters, per-layer complexity O(N·k·e), same order as MHCA with small constant overhead.

Individual. “Precison” is a typo. R1.W5: Eq.(2) uses y_i∈{0,1}^d, element-wise BCE over d then averaged over k; Eq.(4) = per-disease binary softmax over R^{d×2}. R2.W4+C4: STREAM(TMI’25) IS LLM-based (TinyLlama-1.1B), DCAF surpasses it. Direct comparison with LLaVA-Med/MedGemma isn’t apples-to-apples vs ours, would require a separate study under matched compute. SOTA→ARRG-matched.



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 is invited to rebuttal because the reviews indicate a potentially strong and clinically relevant contribution, with generally positive empirical results and supportive ablations, but also several unresolved concerns. The rebuttal should clarify the novelty of the proposed components, experimental validity, especially whether the evaluation protocols and comparisons are fair, important implementation/reproducibility details, and clinical reliability.

  • 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 paper addresses retrieval-augmented radiology report generation by improving disease-consistent retrieval, selective fusion of retrieved text, and disease-aware feature learning. The method is empirically strong on MIMIC-CXR and IU X-Ray, and sufficiently justified for acceptance.



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.

    This paper proposes DCAF, a retrieval-augmented radiology report generation framework that aims to improve disease consistency through disease-aware retrieval, optimal-transport-based cross-modal fusion, and disease-guided feature enhancement. The work reports consistent improvements on MIMIC-CXR and IU X-Ray, with ablations suggesting that each module contributes to the final performance. However, the methodological novelty appears limited for a CXR report generation paper, as most components are incremental adaptations of existing ideas. More importantly, the paper’s central claims about reliable report generation and disease-consistent visual-text alignment are not sufficiently validated: clinical evaluation relies only on CheXbert-based metrics, despite CheXbert-derived labels also being used in the method, and no additional clinical metrics such as RadGraph-F1, GREEN, or RadCliQ are reported. The paper also lacks even lightweight direct evidence for improved visual-text alignment, such as t-SNE visualization or retrieval/alignment analysis, while the computational cost of retrieving from the large training corpus (like MIMIC-CXR) remain insufficiently analyzed. In addition, the comparison across datasets is not fully consistent, with some strong baselines reported on MIMIC-CXR but missing on IU X-Ray.



Meta-review #3

  • 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 most of the concerns raised by reviewers. I recommend Accept.



back to top