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

Abstract

Automated radiology report generation from 3D CT volumes holds immense potential to alleviate radiologist workloads. Current methods, however, face a critical trade-off between efficiency and resolution: many often downsample volumes to fit memory constraints, discarding fine-grained anatomical details essential for diagnosis, while 3D-native models remain computationally prohibitive. To address this, we propose HiFi-Rep, a streamlined framework that leverages high-fidelity visual features without the burden of 3D encoders. Our method processes native 512 x 512 resolution slices across tri-planar views using the pre-trained DINOv3 vision foundation model. Furthermore, to resolve sequence length bottlenecks, we introduce HiFi-S2D, an efficient aggregation module that employs Space-to-Depth transformation. This allows for spatial-preserving aggregation, condensing volumetric depth while strictly retaining high-resolution in-plane diagnostic cues for MedGemma, a specialized medical LLM. We validate HiFi-Rep on the large-scale CT-RATE report generation benchmark, where it establishes a new state-of-the-art. Compared with the best-performing baseline, our proposed approach yields a 9.2% improvement in BLEU-1 and achieves the highest reported METEOR score among existing methods. These results demonstrate that combining robust 2D foundation models with structure-preserving aggregation offers a superior strategy for 3D medical understanding, paving the way for scalable, high-precision automated reporting. Our code is publicly available at https://github.com/rqobistp/HiFi-Rep.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/3274_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/rqobistp/HiFi-Rep

Link to the Dataset(s)

CT-RATE: https://huggingface.co/datasets/ibrahimhamamci/CT-RATE

BibTex

@InProceedings{LimKee_HiFiRep_MICCAI2026,
        author = { Limaroon, Keetawan AND Namwongsa, Punnawat AND Tarnpradab, Sansiri},
        title = { { HiFi-Rep: Leveraging High-Resolution Vision Foundation Models with Tri-Planar Slice Context for CT 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 presents HiFi-Rep, a framework designed for high-resolution 3D CT report generation that uses 2D vision foundation models paired with specialized medical LLMs. The authors address the resolution-efficiency paradox in 3D medical imaging by avoiding computationally expensive 3D encoders and instead processing native 512x512 resolution slices along axial, coronal, and sagittal planes using a frozen DINOv3 backbone. Technical novelties include the HiFi-S2D aggregation module, which combines inter-slice mean pooling with space-to-depth transformations to manage token sequence length, and a tri-planar instruction strategy that uses natural language markers to help the MedGemma LLM align 3D context without explicit 3D positional encodings.

  • 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 framework maintains 512 x 512 inputs rather than downsampling to smaller resolutions, which is critical for the clinical detection of subtle lesions, as well as keeping some 3D context by using axial, coronal, and sagittal information.

    • HiFi-Rep establishes a new state-of-the-art on the CT-RATE benchmark with significant scores in BLEU-1 and METEOR metrics.

    • By using training-free features from DINOv3 and a domain-specific LLM, the model reduces the heavy computational burden typically associated with training 3D-native architectures.

  • 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 authors designate the HiFi-S2D (Space-to-Depth) module as “lossless,” which is technically accurate for the Pixel Shuffle operation itself. However, this is preceded by inter-slice mean pooling along the depth dimension. Mean pooling is a fundamentally lossy operation that collapses 3D volumetric variance into a single 2D average, effectively erasing the “High-Fidelity” (HiFi) depth information before it ever reaches the S2D module. Authors should discuss this limitation instead of relying on the word “lossless”.

    2.The paper uses a frozen DINOv3 (a general-purpose 2D foundation model) while discarding MedGemma’s native MedSigLIP encoder. While 2D foundation models inherit “intelligence” from billions of natural images, they lack domain-specific features for medical imaging. The authors provide no ablation study or direct comparison to prove that DINOv3’s general-purpose features outperform medical-specific encoders (like MedSAM or MedSigLIP) in identifying the subtle textures and low-contrast boundaries unique to CT scans.

    3.A critical finding in the ablation study is that the “Axial Only” configuration achieves a 0.520 BLEU-1 score, whereas adding Coronal and Sagittal views (the full Tri-Planar setup) only increases this to 0.532.This marginal 0.012 gain suggests that the LLM may not be effectively performing tri-planar reasoning. Without a qualitative analysis showing specific clinical cases where a diagnosis was only possible due to the multi-axial views, the increased computational complexity of processing three planes remains poorly justified.

    4.The use of tri-planar natural language markers does not seem optimal, as the model continues to struggle with localization error. This indicates that while the LLM can identify what a pathology is (lexical accuracy), it lacks the native mathematical 3D understanding to accurately place the pathology in 3D space. The absence of explicit 3D positional encodings or cross-view attention mechanisms leaves the model reliant on inferred reasoning, which is prone to spatial hallucinations. Please discuss potential future works or insights to account for this limitation.

  • 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 has provided an anonymized link to the source code, dataset, or any other dependencies.

  • Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?

    N/A

  • Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html

    N/A

  • Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.

    (4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal

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

    This is a solid engineering effort that yields interesting leaderboard results for CT report generation. However, the scientific contribution is weakened by the lossy nature of the aggregation strategy and the marginal improvements gained from the multi-view approach. It would be a much stronger paper if the authors could prove that DINOv3 is fundamentally better than medical-specific encoders through rigorous ablation, and that such model actively benefits from their high-resolution inputs.

  • Reviewer confidence

    Somewhat confident (2)

  • [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.

    Accept

  • [Post rebuttal] Please justify your final decision from above.

    I thank the authors for their thorough and well-structured rebuttal, which has successfully addressed my main concerns. I still maintain that a direct ablation study comparing DINOv3 against MedGemma’s native MedSigLIP encoder would be a valuable avenue for future work to explicitly quantify the benefits of general dense features versus medical-specific encoders. However, this omission does not prevent acceptance, as the current framework is already a solid and complete contribution.



Review #2

  • Please describe the contribution of the paper

    This paper presents a method for generating CT reports by using high-resolution slices across tri-planar views, compressing the visual information using Space-to-Depth transformation, and then using a specialized VLM, MedGemma to generate the report.

  • 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.

    First, the method starts with high-resolution CT instead of heavily down sampling the input at preprocessing step, which helps retain fine grained anatomical and pathological details which is specially important for chest CT. As there could be small malignant lung nodule which could lose its context with heavy downsampling.

    Second, it uses a tri-planar representation by extracting features from axial, coronal, and sagittal views, allowing the model to capture complementary spatial information from different anatomical planes.

    Third, the proposed HiFi-S2D aggregation module, by rearranging spatial pixels to channel dimension, allows to solve long sequence bottleneck.

  • 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.

    First : Lossless compression claim - The method claims to preserve information through HiFi-S2D, but it already applies inter-slice mean pooling beforehand, which is a lossy operation that removes slice-level 3D details.

    Second : Weak justification other than saying it produces high quality dense features using DINOv3 over MedGemma’s own vision encoder. Since MedGemma is already a vision-language model, replacing its visual backbone with a natural image encoder introduces unnecessary complexity and extra feature-alignment challenges unless clearly justified by ablations.

    Computational Complexity : Requiring 16 A100 GPUs limits reproducibility and makes the method impractical for many academic environments, lower-cost or compressed alternatives could be discussed.

    Inappropriate evaluation metric for radiology reports : Claiming state-of-the-art performance using only generic text generation metrics is not convincing for radiology report generation. In this setting, a report may appear fluent and achieve high overlap with the reference while still containing clinically incorrect findings. Without radiology-specific evaluation, the reported gains are difficult to interpret and may overstate the true clinical usefulness of the method. Could use metric like RadGraphF1, GREEN

    1.Ostmeier, Sophie, et al. “Green: Generative radiology report evaluation and error notation.” Findings of the association for computational linguistics: EMNLP 2024.2024.2.https://github.com/Stanford-AIMI/radgraph

  • 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 has provided an anonymized link to the source code, dataset, or any other dependencies.

  • Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?

    N/A

  • Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html

    N/A

  • Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.

    (3) Weak Reject — marginally below the acceptance threshold, but would not mind if accepted, dependent on rebuttal

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

    I suggest following for rebuttal : (1) evaluation should include at least one radiology-specific clinically meaningful metric, not only generic NLG metrics. (2) the paper could tone down the lossless compression claim, since slice information is already lost during inter-slice mean pooling before the Space-to-Depth step. (3) the authors should provide a stronger justification for using DINOv3 instead of MedGemma’s native image encoder, since this choice adds unnecessary complexity unless supported by clear evidence.

  • 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.

    I have updated my rating but I still have several reservations. I encourage the authors to strengthen the validation, particularly by including experiments with the MedSigLIP image encoder. Since the decision to change the encoder is not currently supported by empirical evidence for this specific task, reporting a baseline with the existing image encoder would make the rationale much more convincing. I also strongly recommend including clinically meaningful evaluation metrics; without such evaluation, it is difficult to fully assess the practical value of a radiology-report-oriented method. While I now upgraded rating, I would be cautious about claiming state-of-the-art performance without these additional validations.



Review #3

  • Please describe the contribution of the paper

    This paper presents HiFi-Rep, a novel framework designed for automated radiology report generation from 3D CT volumes. The method addresses the computational bottlenecks and resolution loss typically associated with 3D medical imaging pipelines. This paper proposes parameter-free compression mechanism that uses inter-slice mean pooling followed by a Space-to-Depth operation (Pixel Shuffle). This reduces sequence length for the LLM while losslessly shifting spatial details into the channel dimension. The framework establishes a new state-of-the-art on the large-scale CT-RATE benchmark, achieving a 9.2% improvement in BLEU-1 over the strongest baseline.

  • 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.Bypassing 3D-native models in favor of a frozen 2D foundation model (DINOv3) is computationally astute. By exclusively fine-tuning the linear projector and LoRA adapters on the LLM side, the framework remains highly resource-efficient while still processing native 512x512 slices. 2.The HiFi-S2D module is a well-motivated solution to the resolution-efficiency paradox. Leveraging Space-to-Depth rather than destructive downsampling or linear projection ensures that high-frequency diagnostic details necessary for spotting subtle lesions are preserved. 3.The model demonstrates clear superiority over both volumetric and 2D slice-based methods.

  • Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.

    1.The methodology is exclusively validated on a single benchmark (CT-RATE). While it is a large-scale dataset, evaluating the model’s zero-shot or fine-tuned generalization on an external, out-of-distribution dataset (e.g., CTRG[1]) would greatly strengthen the claims of robustness. 2.As noted by the authors in their qualitative analysis, the model still suffers from hallucinations and omissions in non-target regions. While common in LLM-based generators, minimizing these clinical inaccuracies is paramount for real-world deployment. 3.The paper relies entirely on automated NLP metrics (BLEU, ROUGE, METEOR, BERTScore). While BERTScore helps measure semantic equivalence, automated metrics notoriously fall short in assessing true clinical validity. A reader study or evaluation by board-certified radiologists scoring for clinical correctness (e.g., F1-RadGraph or similar clinical efficacy metrics) is missing.

    References: [1] Tang, Yuhao, et al. “Work like a doctor: Unifying scan localizer and dynamic generator for automated computed tomography report generation.” Expert Systems with Applications 237 (2024): 121442.

  • 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 has provided an anonymized link to the source code, dataset, or any other dependencies.

  • Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?

    N/A

  • Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html

    N/A

  • Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.

    (5) Accept — should be accepted, independent of rebuttal

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

    The proposed HiFi-Rep framework offers an elegant, computationally efficient approach to 3D CT report generation, backed by strong state-of-the-art results on a large-scale benchmark. Although the lack of human clinical evaluation and cross-dataset validation are notable limitations, the paper’s technical novelty and comprehensive ablation studies provide sufficient merit to benefit the MICCAI community.

  • 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.

    N/A



Author Feedback

We sincerely thank the Reviewers for their constructive feedback and for recognizing our framework’s high-resolution preservation and state-of-the-art results on CT-RATE. As per the MICCAI Rebuttal Guidelines prohibiting new experimental results, we address the major concerns using the existing evidence in our manuscript.

1.”Lossless” Claim in HiFi-S2D Aggregation (R1, R2): We agree with R1 and R2 that inter-slice mean pooling collapses depth variance and is inherently lossy. Our “lossless” phrasing specifically referred to the subsequent Space-to-Depth transformation, which shifts pixels to the channel dimension without destructive spatial downsampling. We apologize for this ambiguity. In the revision, we will replace “lossless compression” with “spatial-preserving aggregation” to accurately reflect that while in-plane spatial details are preserved, depth information is condensed.

2.Justification for DINOv3 over Medical Encoders (R1, R2): We thank the reviewers for this suggestion. Our choice of DINOv3 over MedSigLIP is driven by the architectural needs of our framework to preserve high-resolution details. While contrastive models excel at global alignment, they inherently dilute fine-grained spatial features. To effectively capture subtle lesions without destructive downsampling, our HiFi-S2D module strictly requires the highly localized, dense features that DINOv3’s patch-level objective natively provides. Furthermore, recent findings in Raptor [1] empirically confirm that general-purpose DINO models extract richer features for 3D medical volumes than domain-specific encoders (e.g., MedSAM, LLaVA-Med) in train-free settings. We will add a discussion on this architectural trade-off and explore emerging high-resolution medical encoders in our Future Work.

3.Use of Clinical Evaluation Metrics (R2, R4): We fully agree with R2 and R4 that generic NLP metrics are limited in assessing clinical efficacy. Our evaluation strictly followed the NLP metrics established by the CT-RATE benchmark and prior baselines (E3D-GPT, SAMF) to ensure a fair, direct comparison (Table 1). While rebuttal constraints prevent us from providing new RadGraphF1/GREEN scores, we will explicitly incorporate a discussion on the critical necessity of clinical efficacy metrics in our Limitations section.

4.Tri-Planar Marginal Gains & Spatial Hallucinations (R1, R4): While the +1.2% BLEU-1 gain appears numerically modest, it represents a consistent, step-wise improvement across all evaluation metrics as views are incrementally added, achieving strict statistical significance across the board (paired t-test, p < 0.01). Furthermore, this inherently mimics the standard radiological workflow, where coronal/sagittal reformations are essential to resolve spatial ambiguities. In our revision, we will add qualitative case comparisons demonstrating where “Axial Only” fails but Tri-Planar succeeds. Regarding spatial hallucinations, we explicitly acknowledged this limitation in our Qualitative Analysis (Fig 2).

5.Hardware Requirements and Reproducibility (R2): The 16 A100 GPUs were used strictly to accelerate DDP training on the massive 47k CT-RATE dataset, not due to VRAM constraints. Because DINOv3 is frozen and we exclusively employ LoRA (r=64), the memory footprint remains inherently minimal. While we intentionally adopted a standard training protocol to rigorously validate HiFi-Rep’s fundamental efficacy, total training took just ~2 days (16.5h for alignment, 34.5h for instruction tuning). To ensure reproducibility, our released codebase will provide optimized Unsloth scripts, enabling seamless replication on a single consumer GPU (e.g., 24GB/40GB).



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.

    This paper received three good quality reviews with two accept and one reject rating. The final score will be depended on the rebuttal quality and details.

    The paper presents a high-resolution tri-planar CT report generation framework that uses frozen DINOv3 features, HiFi-S2D aggregation, and LLM adaptation to preserve 512×512 detail while reducing the cost of 3D-native modeling, achieving strong CT-RATE results on generic text metrics. Reviewers appreciated the engineering design, computational efficiency relative to full 3D models, and state-of-the-art benchmark performance, but questioned the lossless claim because inter-slice mean pooling is lossy, the limited gain from tri-planar views, the lack of proof that DINOv3 outperforms medical encoders, and the high resource requirement of 16 A100 GPUs. The missing pieces are: ablations comparing DINOv3 with medical vision encoders and axial-only versus tri-planar reasoning, clinically meaningful report metrics such as RadGraphF1/GREEN or radiologist evaluation, hallucinations/localization errors and cases where multi-view input helps, and generalization with external CT report datasets beyond CT-RATE.

  • 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.

    After rebuttal, all three reviewers agreed on recommending acceptance for this paper. They all liked the technical simplicity, novelty part. While it is not reqired for the final version, AC would consider the following is quite reasonable and this is not fully addressed by the rebuttal. Thus the clinical indications or impacts for this paper is still unclear.

    “Inappropriate evaluation metric for radiology reports : Claiming state-of-the-art performance using only generic text generation metrics is not convincing for radiology report generation. In this setting, a report may appear fluent and achieve high overlap with the reference while still containing clinically incorrect findings. Without radiology-specific evaluation, the reported gains are difficult to interpret and may overstate the true clinical usefulness of the method. Could use metric like RadGraphF1, GREEN

    1.Ostmeier, Sophie, et al. “Green: Generative radiology report evaluation and error notation.” Findings of the association for computational linguistics: EMNLP 2024.2024.2.https://github.com/Stanford-AIMI/radgraph”



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.

    There are 3 Accepts after rebuttal. Reviewers generally found the framework useful and well engineered, with strong results on CT-RATE. The rebuttal clarified the lossy aggregation issue, hardware requirements, and the motivation for using DINOv3.Some concerns remain, especially the lack of clinical report metrics, external validation, and direct comparison with medical vision encoders. However, the reviewers were sufficiently satisfied after rebuttal, and the work makes a solid practical contribution. I recommend acceptance.



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.

    This paper initially received two accept and one reject rating. After rebuttal, authors successfully address reviewers concerns, and all reviewers agree for acceptance. After reading the paper, all review comments, I agree that the acceptance of this paper.



Meta-review #4

  • 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.

    This paper received three reviews with initial scores ranging from Weak Reject to Accept, and after the rebuttal, all three reviewers updated to Accept. The authors’ rebuttal addressed the major concerns raised by the reviewers. Regarding the “lossless” compression claim, the authors acknowledge that inter-slice mean pooling is inherently lossy and propose rephrasing it as “spatial-preserving aggregation.” For the choice of DINOv3 over medical-specific encoders, the authors provided architectural justification (need for dense, patch-level features for HiFi-S2D) and cited supporting literature (Raptor), though Reviewer #2 still noted the absence of direct empirical comparison with MedSigLIP as a remaining limitation. On clinical evaluation metrics, the authors agreed with reviewers but explained their adherence to CT-RATE benchmark standards for fair comparison, promising to add a discussion of clinical metrics in the limitations. The marginal tri-planar gains were defended with statistical significance (p<0.01) and workflow justification, with promised qualitative cases. Overall, the rebuttal was thorough and responsive, successfully converting all reviewers to acceptance despite some remaining reservations. It will be an interesting contribution to the community.



back to top