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

Abstract

Automatic chest X-ray report generation requires models to capture fine-grained visual evidence and describe findings within clinical context. Existing approaches often entangle global structure and local details, making subtle cues prone to attenuation during compression and cross-modal alignment. Most frameworks also rely on single-image inputs and underuse patient priors, limiting their ability to reflect diagnostic intent and disease evolution. We propose Wavelet-Enhanced Radiology Report Generation (WERG), a clinically motivated coarse-to-fine framework built on a multimodal large language model that supports joint conditioning on clinical context and prior images. WERG generates an initial draft, retrieves observation-aligned report snippets as textual priors, and refines the report with enhanced visual evidence. Moreover, we introduce a wavelet-based frequency-domain detail enhancement module to emphasize subtle abnormalities while preserving global semantics, and a diagnosis-guided calibration branch trained with weak observation labels. Experiments on MIMIC-CXR and IU X-Ray show consistent gains on NLG and clinical metrics, producing more detailed and clinically consistent reports.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/1545_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/CodeCQ789/WERG

Link to the Dataset(s)

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

BibTex

@InProceedings{YanShu_WaveletEnhanced_MICCAI2026,
        author = { Yang, Shuanghong AND Deng, Yan AND Qin, Qibing AND Zeng, Pan AND Zhang, Henggui AND Hu, Wei AND Zhang, Wenfeng},
        title = { { Wavelet-Enhanced Coarse-to-Fine Radiology Report Generation with Large Language Models } },
        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

    This paper introduces WERG, a framework for radiology report generation that utilizes wavelet transforms to enhance high-frequency visual details and a retrieval mechanism to inject clinical priors. The method aims to solve the problem of subtle detail loss during feature compression. Experiments show it achieves state-of-the-art performance on the MIMIC-CXR dataset in both language generation and clinical accuracy 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 use of wavelet transforms to explicitly separate and amplify high-frequency visual details is a technically solution to the common problem of “detail attenuation” in medical image analysis. (2) The combination of visual enhancement with the RRKE retrieval module creates a strong coarse-to-fine generation pipeline that logically aligns with clinical diagnostic reasoning. (3) The results are impressive, setting a new benchmark on the MIMIC-CXR dataset with significant gains in clinically relevant metrics.

  • 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 “Related Work” is insufficient. The introduction fails to adequately contrast WERG with recent LLM-based medical generators. Please explicitly discuss how your wavelet method differs from other visual enhancement techniques in recent literature. (2) The ablation study is confusing. Since DGVFC depends on FDE, testing “FDE alone” is misleading. Please restructure the ablation to test the combined visual enhancement block (FDE+DGVFC) against the baseline for clarity. (3) The IU X-Ray evaluation is weak. Testing in a zero-shot setting without fine-tuning makes comparison difficult. Please either fine-tune on IU X-Ray or provide a stronger justification for the zero-shot approach. (4) Figure 2 lacks visual proof. The qualitative examples need bounding boxes or zoomed insets on the X-rays to prove the model is actually detecting the “subtle cues” mentioned in the text.

  • 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

    There are grammatical issues. For example, “phrase findings under clinical context” in the abstract is awkward. A thorough language edit is recommended.

  • 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 manuscript requires significant revisions to strengthen the ablation studies, clarify the methodological dependencies, and improve the presentation of qualitative results before it can be considered for publication.

  • 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.

    After reading the rebuttal, I think the authors addressed most of the major concerns reasonably well. In particular, they clarified the motivation and role of the FDE+DGVFC modules, explained the ablation design more clearly, and acknowledged the issue with inconsistent clinical metrics in the comparison table. The response on the IU X-Ray zero-shot setting was also reasonable and consistent with prior work.



Review #2

  • Please describe the contribution of the paper

    This paper focuses on the automatic report generation of the chest X-ray. It proposes a coarse-to-fine framework WERG which is built on the multimodal LLM. In the framework, it combines four important parts: retrieval-augmented text priors (RRKE), wavelet-based frequency-domain visual enhancement (FDE), and diagnosis-guided feature calibration (DGVFC). Most of the experiments on MIMIC-CXR and IU X-Ray show consistent and relatively improved performance, while gains on n-gram overlap metrics such as BLEU-1 and ROUGE-L are relatively modest.

  • 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.Clinical Relavance: This paper focus on the automatic generation of the X-ray report. The framework it proposes tries to generate finer reports with consistency and accuracy. And this application can truly reduce the radiologists’ documentation burden and make the radiology workflow more consistent. 2.Methodological novelty: The framework it proposes is different from the one-pass methods of LLM generation in the past. It adds three components in the architecture: report retrieval, wavelet-based frequency-domain detail enhancement, and diagnosis-guided visual feature calibration. 3.Experimental effort: The paper finishes experiments on two public chest X-ray benchmarks, MIMIC-CXR and IU X-Ray. The baselines /competitive models it chooses are all reasonable and proper. The ablation study design is also complete to prove every new components’ contributions.

  • 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 paper is of generally good quality, but several aspects would benefit from further clarification and discussion. 1.The combination of RRKE, FDE, DGVFC is reasonable, but the components are not particularly novel in isolation. Wavelet decomposition for feature enhancement has been explored in visual representation learning (like Wave-ViT, ECCV 2022), and retrieval-augmented report generation has been studied in the work like PromptMRG (AAAI 2024). The contribution therefore lies primarily in their integration, which could be positioned more explicitly and justified more carefully. 2.The cross-dataset results on IU X-Ray are a little bit mixed. WERG achieves the best MTR and R-L but falls below the strongest baselines on B-4, and the authors’ explanation that prior-guided refinement introduces paraphrasing and sentence reordering is qualitative and not empirically verified. This makes the generalization benefit less convincing, especially since IU X-Ray is used in a zero-shot setting and only a limited set of metrics is reported there. A more systematic analysis of when and why the refinement stage degrades n-gram precision would strengthen the paper. 3.Clinical Usability: The sensitivity of retrieval quality to the size and domain of this database, as well as potential data privacy concerns in clinical settings, are not addressed. (But this point is relatively less important in the weakness considering the limitation of the paper pages.)

  • 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?

    I think this paper is relevant and technically reasonable, especially it tries to integrate wavelet-based visual enhancement with retrieval-augmented generation into a multimodal LLM. The experiment results on MIMIC-CXR are good, and the ablation study provides evidence for the proposed modules. However, my overall score is mainly limited by concerns regarding the limited novelty of individual components, the mixed performance on IU X-Ray. But I still look forward to authors’ feedback and would like to give higher scores if I can receive more detailed explanation from the authors.

  • 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.

    I support acceptance. The paper is technically reasonable and clinically relevant, although some concerns remain. In the final version, the authors should make the novelty claim more modest, discuss the IU X-Ray metric inconsistency more carefully, and add a clearer limitation/privacy discussion for the retrieval module.



Review #3

  • Please describe the contribution of the paper

    This paper proposes WERG, a coarse-to-fine multimodal-LLM framework for chest X-ray report generation. The model conditions on the current CXR, optional clinical context, and prior images, then refines an initial draft using retrieved report priors (RRKE), a wavelet-based frequency-domain detail enhancement module (FDE), and diagnosis-guided visual feature calibration (DGVFC). The main intended contribution is to improve fine-grained finding coverage and clinical consistency, especially for subtle abnormalities and comparison-aware reporting.

  • 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.

    S1.Clinically meaningful problem setting. The paper addresses a clinically meaningful problem setting by incorporating the current image, optional clinical context, and prior studies, which are relevant for comparison-aware chest X-ray reporting. S2.The main technical novelty is in FDE/DGVFC. The most distinctive part of the paper is the visual-side design in Sec. 2.3–2.4: frequency-domain detail enhancement (FDE) and diagnosis-guided visual feature calibration (DGVFC). This is the clearest novel aspect of the submission. S3.The paper includes a reasonably complete evaluation package. The submission reports results on MIMIC-CXR and zero-shot IU X-Ray, includes ablations, and provides qualitative examples.

  • 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.

    W1.The paper has some novelty in FDE/DGVFC, but the overall methodological novelty is limited. From Sec. 1 and Sec. 2.1–2.2, WERG is presented as a broad coarse-to-fine, diagnosis-aware, retrieval-augmented multimodal-LLM framework, but the paper does not clearly separate inherited ideas from the actual new contribution. Diagnosis-aware prompting already appears in PromptMRG [1], retrieval/knowledge-augmented LLM report generation is already explored in KACL [2] and RADAR [3], and knowledge-enhanced LLM report generation is already explored in KARGEN [4]. After a strict reading, the part that appears genuinely new is mainly FDE/DGVFC, and that novelty does not feel strong enough to support the broader novelty claim of the full pipeline. Also, RADAR [3] is cited in the introduction but omitted from the comparison table despite being one of the closest baselines. W2.The experimental comparison could be strengthened; the comparison set is incomplete, the criteria for ensuring a fair evaluation are not explicitly defined, and some clinical metrics appear non-comparable. WERG’s Table 1 presents its MIMIC-CXR METEOR = 0.347 as the highest among the listed baselines, but the paper does not explain that several strong recent MLLM baselines with substantially higher METEOR are omitted. In RADAR [3], the MIMIC-CXR comparison includes LLaVA-Med [5] (0.353), MAIRA-1 [6] (0.333), MAIRA-2 [7] (0.420), Libra [8] (0.489), and RADAR [3] itself (0.450). Without discussing such baselines, the current table gives an incomplete picture of the recent literature. In addition, the paper does not state that baselines were re-run under a unified protocol, so preprocessing, report-section choice, tokenization, and evaluation details may differ across papers. A further concern is the KARGEN 0.203 entry: KARGEN reports 0.203 as RadGraph F1 [9], whereas WERG labels the clinical columns as CheXbert-style Precision/Recall/F1 [10]. Ultimately, the use of disparate metrics in the same comparison prevents a fair assessment of the results, thereby limiting the overall credibility and interpretability of the clinical validation section. W3.The ablation study is not strong enough to support the claimed contribution. In Table 3, DGVFC is only evaluated together with FDE, so its individual contribution is not isolated. There is also no comparison against a simpler enhancement baseline. As a result, the paper shows that the full stack helps, but not that the specific wavelet-based design is necessary. W4.Reproducibility is weak. The paper does not provide code, and the method description is not detailed enough for reliable reproduction. In Sec. 2.1–2.4, the exact instantiated vision backbone/LLM is not clearly specified, and key details remain vague, including the retrieval classifier g(⋅), the label space used in RRKE/DGVFC, the rule for “relatively less emphasized” sentence selection, and the training/loss details for the full model. This makes the results difficult to audit and hard to reproduce. References [1] 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). [2] Sha, Y., Pan, H., Meng, W., Li, K.: Contrastive Knowledge-Guided Large Language Models for Medical Report Generation. In: Medical Image Computing and Computer Assisted Intervention – MICCAI 2025, LNCS, pp. 111–120.Springer, Cham (2025). [3] Hou, W., Cheng, Y., Xu, K., Li, H., Hu, Y., Li, W., Liu, J.: RADAR: Enhancing Radiology Report Generation with Supplementary Knowledge Injection. In: Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics, pp. 26366–26381 (2025). [4] Li, Y., Wang, Z., Liu, Y., Wang, L., Liu, L., Zhou, L.: KARGEN: Knowledge-enhanced Automated Radiology Report Generation Using Large Language Models. In: Medical Image Computing and Computer Assisted Intervention – MICCAI 2024, LNCS, vol. 15005, pp. 382–392.Springer, Cham (2024). [5] Li, C., Wong, C., Zhang, S., Usuyama, N., Liu, H., Yang, J., Naumann, T., Poon, H., Gao, J.: LLaVA-Med: Training a Large Language-and-Vision Assistant for Biomedicine in One Day. In: NeurIPS Datasets and Benchmarks (2023). [6] Hyland, S.L., Bannur, S., Bouzid, K., Castro, D.C., et al.: MAIRA-1: A Specialised Large Multimodal Model for Radiology Report Generation. arXiv preprint arXiv:2311.13668 (2023). [7] Bannur, S., Bouzid, K., Castro, D.C., et al.: MAIRA-2: Grounded Radiology Report Generation. arXiv preprint arXiv:2406.04449 (2024). [8] Zhang, X., Meng, Z., Lever, J., Ho, E.S.L.: Libra: Leveraging Temporal Images for Biomedical Radiology Analysis. Findings of ACL, pp. 17275–17303 (2025). [9] Jain, S., Agrawal, A., Saporta, A., Truong, S.Q.H., Duong, D.N., Bui, T., Chambon, P., Zhang, Y., Lungren, M.P., Ng, A.Y., Langlotz, C.P., Rajpurkar, P.: RadGraph: Extracting Clinical Entities and Relations from Radiology Reports. In: NeurIPS Datasets and Benchmarks (2021). [10] Smit, A., Jain, S., Rajpurkar, P., Pareek, A., Ng, A.Y., Lungren, M.P.: Combining Automatic Labelers and Expert Annotations for Accurate Radiology Report Labeling Using BERT. In: Proceedings of EMNLP, pp. 1500–1519 (2020).

  • 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

    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.

    (2) Reject — should be rejected, independent of rebuttal

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    While the paper introduces some technical novelty via FDE/DGVFC, the overall contribution feels thin compared to recent work, and the empirical results aren’t robust enough to support acceptance. 1.Limited Novelty The high-level pipeline is very similar to existing diagnosis-aware and retrieval-augmented frameworks like PromptMRG, KACL, RADAR, and KARGEN. Once you look past the standard coarse-to-fine MLLM setup, the only truly new component is the visual enhancement/calibration module. 2.Experimental Flaws

    • Missing Baselines: The authors claim a METEOR advantage but ignore several top-tier MLLM baselines (like LLaVA-Med, MAIRA-1/2, and Libra) that actually report much higher scores on MIMIC-CXR.
    • Unfair Comparison: There’s no mention of a unified protocol for the baselines. Without knowing if the preprocessing, tokenization, and metric computation were standardized, the comparison isn’t reliable.
    • Metric Confusion: The clinical table seems to mix apples and oranges—listing RadGraph F1 for KARGEN alongside CheXbert metrics for WERG. These aren’t interchangeable. 3.Reproducibility The paper is quite hard to follow from an implementation standpoint. There’s no code, the specific LLM backbone isn’t named, and key details—like the retrieval classifier, the RRKE/DGVFC label space, and the specific loss functions—are left vague. Because the novelty is limited and the evaluation is inconsistent, I don’t believe the paper is ready for publication in its current state.
  • 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



Author Feedback

We thank the Meta-Reviewer (MetaR) and all reviewers (R) for their constructive feedback. We respond below organized by key themes. 1.Contribution & Methodological Novelty (R2W1, R3W1) WERG formulates radiology report refinement as jointly constrained by textual priors and fine-grained visual evidence. Existing methods provide textual guidance but lack precise visual support. WERG unifies textual priors, frequency-domain visual enhancement, and diagnostic semantic calibration, enabling observation-level priors and enhanced visual cues to jointly guide report refinement. We introduce wavelet-based enhancement (FDE) to explicitly separate low-frequency structure and high-frequency details, enhancing boundaries, textures, and local density cues before LLM decoding. DGVFC aligns these enhanced features with diagnostic semantics, improving grounding for report refinement. This integration addresses weak synergy in prior RRG pipelines, forming a targeted, clinically relevant solution rather than a simple stack of modules. 2.Ablation & Module Justification (R1W2, R3W3, R1W1) Ablation analyses clarify each module’s contribution. Table 3 compares the baseline, FDE-alone, and FDE+DGVFC. FDE+DGVFC consistently improves BLEU-4/METEOR/ROUGE-L (+0.008/+0.013/+0.003) and CE Precision/Recall/F1 (+0.010/+0.128/+0.075) over the baseline. FDE-alone serves as a control to ensure wavelet enhancement preserves base visual representation. Compared to FDE-alone, DGVFC further improves recall by 0.108 without reducing precision, indicating enhanced coverage of positive observations. Wavelet decomposition is necessary: CXR report generation requires both global structure and high-frequency details, which are more controllably enhanced than with unconstrained feature learning. FDE is not generic frequency decomposition: high-frequency sub-bands are fused as residuals and low-frequency components modulate channels, preserving global semantics while enhancing fine-grained evidence for LLM refinement. 3.Experimental Protocol & Dataset Settings (MetaR, R1W3, R2W2, R3W1, R3W2) We follow prior work for zero-shot IU X-Ray evaluation (PromptMRG AAAI’24; DDaTR TMI’25), noting that the standard IU test split may be unsuitable for disease-aware metrics. Comparisons use example-based CE; consistent with top-tier papers (HSA MICCAI’24; DCL CVPR’23), we cite and fairly compare recent methods. For KARGEN, we acknowledge an oversight and will remove its clinical entry while keeping NLG metrics. Reported IU X-Ray performance differences (e.g., lower BLEU-4) can arise from wording variability despite correct diagnostic predictions, as noted in prior work. Higher METEOR/ROUGE-L indicates improved semantic coverage and robustness to distribution shift.Missing comparisons (MAIRA-1/2, Libra) were excluded due to clear differences in model scale(7B vs our 3.8B). RADAR reproduction under controlled settings did not outperform WERG, making direct inclusion unfair. 4.Clinical Significance & Qualitative Visualization (R3W1, R1W4) WERG targets real intrapartum CXR scenarios, supporting observation-level report refinement and clinically relevant measurements. We will enhance Figure 2 with bounding boxes to improve interpretability. The method’s alignment of textual priors and fine-grained visual evidence ensures outputs are semantically and diagnostically coherent, supporting clinical applicability. 5.Reproducibility, Related Work, & Language (R3W4, R1W1, R1) We will release full implementation details, code, and model weights upon acceptance to facilitate reproducibility. Due to space, not all recent LLM-based RRG models are compared in the main paper, but we will include discussion of LLaVA-Med, MAIRA-1/2, and Libra in the final version. We will thoroughly polish the manuscript language for clarity. 6.Data Privacy & Retrieval (R2W3) We acknowledge the concern regarding retrieval and data privacy and will add discussion of privacy-aware retrieval in the final version.



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 reviewers recognize the clinical importance of automating radiology report generation and find the technical approach—specifically the use of wavelet transforms for detail enhancement—to be a promising direction. However, there are significant concerns regarding the fairness of the experimental comparisons and the clarity of the paper’s specific contributions.

    The authors must provide direct and data-driven responses to the following points to avoid a rejection:

    Completeness of Literature and Baselines: Reviewers noted that several recent, high-performing models (e.g., RADAR, MAIRA-1/2, Libra) were omitted from Table 1.You must explain these omissions or update your comparison to include these state-of-the-art results to prove the superiority of WERG.

    Inconsistent Metrics: Clarify the clinical evaluation metrics. Reviewer #3 pointed out a potential “apples-to-oranges” comparison where CheXbert metrics were compared against RadGraph F1 scores from other papers. Please provide a unified metric comparison.

    IU X-Ray Evaluation: Address why the IU X-Ray results were presented in a zero-shot setting and why certain n-gram metrics (like B-4) dropped compared to baselines.

  • 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 clearly clarified the issues raised by the reviewers. The explanations in the rebuttal look reasonable and correct to me.



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 reviewers agree that the paper has clinical relevance and technical merit, with promising integration of wavelet-based visual enhancement and retrieval-augmented generation in multimodal LLMs. Some concerns were raised about the limited novelty of individual components, experimental inconsistencies, and insufficient implementation details, which are addressed in the rebuttal.



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 paper is recommended for acceptance after considering the rebuttal and discussion among Reviewers.

    The rebuttal successfully addressed several major concerns, particularly regarding the metric inconsistency and experimental settings, which improved reviewer confidence in the results. While some questions remain regarding the degree of methodological novelty, the consensus is that these concerns are not sufficient to outweigh the paper’s contributions and empirical validation.

    For the final version, the authors are encouraged to moderate the novelty claims, provide a clearer discussion of the IU X-Ray metric issue, and expand the discussion of limitations, particularly regarding the retrieval module and potential privacy considerations.

    I recommend Accept. I also suggest that the authors revise the novelty discussion and limitations section in the camera-ready version.



back to top