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

Abstract

Ultra-Wide-Field (UWF) retinal imaging has revolutionized retinal diagnostics by providing a comprehensive view of the retina. However, it often suffers from quality-degrading factors such as blurring and uneven illumination, which obscure fine details and mask pathological information. While numerous retinal image enhancement methods have been proposed for other fundus imageries, they often fail to address the unique requirements in UWF, particularly the need to preserve pathological details. In this paper, we propose a novel frequency-aware self-supervised learning method for UWF image enhancement. It incorporates frequency-decoupled image deblurring and Retinex-guided illumination compensation modules. An asymmetric channel integration operation is introduced in the former module, so as to combine global and local views by leveraging high- and low-frequency information, ensuring the preservation of fine and broader structural details. In addition, a color preservation unit is proposed in the latter Retinex-based module, to provide multi-scale spatial and frequency information, enabling accurate illumination estimation and correction. Experimental results demonstrate that the proposed work not only enhances visualization quality but also improves disease diagnosis performance by restoring and correcting fine local details and uneven intensity. To the best of our knowledge, this work is the first attempt for UWF image enhancement, offering a robust and clinically valuable tool for improving retinal disease management.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2025/paper/2387_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

N/A

Link to the Dataset(s)

N/A

BibTex

@InProceedings{LiaWei_AFrequencyAware_MICCAI2025,
        author = { Liao, Weicheng and Chen, Zan and Xie, Jianyang and Zheng, Yalin and Ma, Yuhui and Zhao, Yitian},
        title = { { A Frequency-Aware Self-Supervised Learning for Ultra-Wide-Field Image Enhancement } },
        booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2025},
        year = {2025},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 15972},
        month = {September},
        page = {2 -- 12}
}

Reviews

Review #1

  • Please describe the contribution of the paper

    This paper proposed a frequency-aware self-supervised learning method in addressing the challenges of blurring and uneven illumination in UWF images. The design of the FRED and RICE modules is unique and effectively solves the specific problems of UWF images.

  • 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.This article provides a detailed description of the design and implementation of the FRED and RICE modules, effectively addressing specific issues related to UWF images. 2.The paper combines technologies from the fields of deep learning and signal processing, addressing the image enhancement problem through a frequency-spatial joint optimization framework.

  • 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 comparative experiments are relatively limited, and there are currently many self-supervised SOTA models for dark light enhancement, but none of them have been used in this article. 2.The robustness of the FRED module in dealing with complex blurring situations (such as multiple blurring or extreme blurring) needs further verification. 3.Although frequency domain analysis performs well in deblurring, its ability to handle high-frequency noise and preserve details needs to be demonstrated
  • 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.

  • 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/2025/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 — could be rejected, dependent on rebuttal

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

    This paper proposed a frequency-aware self-supervised learning method in addressing the challenges of blurring and uneven illumination in UWF images. However, the comparative experiments are insufficient, lacking demonstrations of complex fuzzy situations and the ability to preserve high-frequency noise and details.

  • 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.

    This paper proposed a frequency-aware self-supervised learning method in addressing the challenges of blurring and uneven illumination in UWF images. However, the comparative experiments are insufficient, lacking demonstrations of complex fuzzy situations and the ability to preserve high-frequency noise and details.



Review #2

  • Please describe the contribution of the paper

    This paper proposed a new application - Ultra-wide fundus image enhancement - and a new method dedicated to address this problem. The methods used several techniques: (1) decompose the high-freq and low-freq images, (2) retinex-guided illumination compenstation to get illumination components, (3) architecture changes such as ACI and CPU using wavelet transform. The results showed improved performance on both quantitative and qualitative metrics. The ablation study shows the effectiveness of the proposed modules. The authors claimed to be the first working in this application.

  • 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 results are in general promising and self-contained.
    2. The effectiveness of each proposed module is clearly ablated.
    3. The paper is well written in most of the parts and looks self-contained.
  • 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.

    I have some questions to the authors:

    1. The components that bring more contribution seems to be the FRED and RICE framework. I am however not sure how novel they are. Can the author justify they are really novel from existing works?

    2. I am a little bit confused about the implementation of the overall architecture:

    What is the S_1 in Fig. 2a? I can not find caption.

    The color is bullied in 2(a) and 2(b) ACI, the green to purple is used for high-res and low-res separation, but then used with global / local information too, which I don’t understand. I also don’t understand in the first sentence of ACI section, what is the actual composition of C? Are features from all Unet blocks being concatenated together? How was the concatenation and rescale implemented? The author should clarify this.

    Despite I know where APS is applied, I can not find it in Fig. 2. Also, the author should show what will be looking like for the low-freq and high-freq images for UWF images if they claimed this is the first work on it.

    Where is the Feature Fusion module coming from? There is no reference and no introduction.

    1. Maybe I am wrong, but this paper does not mention anything about open sourcing the code or the data or the model. Consider this is the first work on this application, making everything private will hurt the reliability of the work. This is one of my major concern.

    2. Fig. 2: images are are -> images are

    3. Data distribution: “In our experiments, 400 images with varying exposure levels were selected for training, while 434 images with different levels of blur and illumination were used to evaluate the performance of the image enhancement method” I don’t understand what is the difference between exposure level and illumination. Why is the author bringing in such description? Again, not publishing the data makes me worried about the reliability of the training process.

    4. ” Nevertheless, these approaches rely heavily on the availability of extensive collections of high- quality images free from quality-degrading factors.” I don’t think this method solves this problem, in my understanding, the authors also did image synthesis to construct pairwise training data.

    5. I suggest not use ‘CPU’ as your module terminology. I think people might misunderstand this CPU to be computing process unit - the core in the modern computer, which is not good.

  • 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 provide sufficient information for reproducibility.

  • 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/2025/en/REVIEWER-GUIDELINES.html

    I would love to increase score if the authors can:

    1. address my questions about the method and experiment section, and think about how to improve the demonstration.
    2. redraft the introduction to avoid overclaiming the challenges being solved.

    I might decrease my score if the authors fail to:

    1. Don’t give a reasonable response for the reproducibility, including the plan for releasing the data, code and model.
  • 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 — could be accepted, dependent on rebuttal

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

    First, this paper is self-contained. Second, the results looks promising. However, I think the demonstration can be improved. Also, my major worry is the lack of guarantee to the reproducibility of this study, especially if the authors think they are the first working on this problem.

  • Reviewer confidence

    Confident but not absolutely certain (3)

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

    N/A

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

    N/A



Review #3

  • Please describe the contribution of the paper

    The study proposes a novel self-supervised learning method to enhance ultra-wide-field images. The method includes frequency-decoupled image deblurring, asymmetric channel integration, and retinex-based illumination compensation modules.

  • Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.

    The paper clearly describe the proposed modules and their functions in the enhancement process. The paper also compares the proposed method qualitatively and quantitatively with other methods.

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

    The paper should emphasize and discuss more about the effectiveness of the proposed method compared to other methods with more information about their approaches. Advantages and disadvantages of the proposed method should be further discussed. The method’s limitation should also be discussed to enhance the quality of the paper.

  • 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.

  • 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/2025/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 — could be accepted, dependent on rebuttal

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

    The description of the method is clear. Nevertheless, more detailed discussion and analysis of the proposed method compared to other method should be included, so that the novelty and the contribution of the work can be more emphasized.

  • 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 rebuttal addresses my stated concerns in the previous version



Author Feedback

  1. No comparison with self-supervised models (R1) We only compared with supervised and unsupervised methods for fairness, as our framework combines FRED (supervised deblurring) and RICE (unsupervised low-light enhancement). It is worth noting that RICE operates without supervisory signals from unlabeled data, we exclude comparisons with self-supervised in our paper (we did the experiments), even though our work outperforms most self-supervised works.

  2. The robustness on complex blurring (R1) In fact, our evaluation used 434 test images spanning diverse blur conditions, including complex multi-blur and extreme cases, to assess robustness. Results show FRED outperforms MRDNet by 16.8% in BRISQUE, with consistent gains in NIQE/PIQE (Table 1). This strength derives from its multi-scale frequency-band decomposition, which selectively processes distinct blur patterns in their optimal spectral domains.

  3. Ability of handling high-frequency noise (R1) While our paper didn’t include visual examples due to page constraints, it’s important to emphasize that our test dataset rigorously evaluates performance across multi-level blur, illumination changes, and high-frequency noise. Results demonstrate significant superiority over denoising-inclusive methods (URRN-Net, RUAS, ZeroIG), with ≥20.8% BRISQUE improvement. This stems from two innovations: (1) The ACI mechanism merges global noise analysis with local detail preservation, enabling simultaneous noise suppression and structure retention, and (2) Our frequency-aware architecture processes degradations in their optimal domains, avoiding limitations of separate denoising/deblurring approaches.

  4. Novelty and limitation (R2, R3) The FRED and RICE are the major innovations. The former introduces a symmetric architecture that separately processes high/low-frequency information, treating blur as both high-frequency detail loss and low-frequency diffusion, providing more comprehensive frequency-domain restoration than MRDNet that miss fine structures. RICE introduces CPU based on multi-scale decoupling of wavelet and estimates illumination compensation ratios, respectively avoiding color distortion and error propagation of Retinex-based methods like ZeroIG. Limitations include occasional over-exposure (e.g., optic disc regions) due to restricted receptive fields in low-frequency processing. Future work will enhance global illumination modeling in low-frequency domains and expand validation to downstream tasks like vessel quantification.

  5. Detailed issues and reproducibility of our method. (R2) We’ll include missing details, and make the dataset and code publicly accessible. – S_1 in Fig. 2a is the output of FFM1 in Fig. 2b. – Issues in ACI: FRED contains high frequency (green) and high frequency (purple) pathways, both of which adopt U-Net as backbone. In each pathway, ACI takes the features from all the encoder layers firstly by bilinear interpolation to be rescaled to the feature size of the corresponding decoder layer, and then concatenates them along the channel dimensions to get the fused features with channel number C. – FFM implementation follows Ref. 22 (Page 5).

  6. Difference between exposure level and illumination (R2) We apologize for the confusion, in this work, “exposure” and “illumination” are used interchangeably. We will consistently use “illumination” in the final revision.

  7. Overclaiming the challenges being solved (R2) We will redraft the introduction as: “While promising, these methods typically require high-quality image datasets with minimal degradation. In practice, UWF images usually contain both blur and illumination variations, making pristine samples scarce. Our approach uses synthesized training pairs from images with relatively clear structures—not perfect quality—demonstrating practical applicability without needing ideal training data.”



Meta-Review

Meta-review #1

  • Your recommendation

    Invite for Rebuttal

  • If your recommendation is “Provisional Reject”, then summarize the factors that went into this decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. You do not need to provide a justification for a recommendation of “Provisional Accept” or “Invite for Rebuttal”.

    N/A

  • 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. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’

    N/A



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. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’

    N/A



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. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’

    Please address the comments in reviews, like “visualizing the high-freq / low-freq features”.



back to top