List of Papers Browse by Subject Areas Author List
Abstract
Low-quality fundus images are pervasive in real-world clinical screening, severely compromising the reliability of diabetic retinopathy (DR) diagnosis and degrading the performance of automated grading systems. Existing methods either rely on scarce paired training data or treat enhancement as a uniform image-level operation, neglecting the spatially varying importance of clinically sensitive regions such as lesions and vessels. To address these limitations, we propose a semi-supervised dual framework that integrates unsupervised knowledge distillation-driven quality transfer with importance-guided supervised enhancement. The quality transfer module aligns the global quality distribution of low-quality images to the high-quality domain by exploiting the representational discrepancy between source and target CNN models, without requiring paired training data. The importance-guided module employs a pixel-level importance decoder to adaptively weight the supervision signal, prioritizing clinically critical regions during reconstruction. Extensive experiments on the EyeQ and Messidor benchmarks demonstrate that our method achieves state-of-the-art enhancement performance, outperforming all compared methods. Moreover, the enhanced images yield significant improvements in downstream DR grading, validating the clinical utility of our framework.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/2528_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
Link to the Dataset(s)
DDR dataset: https://github.com/nkicsl/DDR-dataset
EyePACS dataset: https://www.kaggle.com/c/diabetic-retinopathy-detection/data
EyeQ dataset: https://github.com/HzFu/EyeQ
Messidor dataset: https://www.adcis.net/en/third-party/messidor/
BibTex
@InProceedings{HouQin_Knowledge_MICCAI2026,
author = { Hou, Qingshan AND Yang, Wenju AND Wang, Meng AND Xue, Peiyun AND Huang, Xuhui AND Liu, Yangyang},
title = { { Knowledge Distillation-Driven Quality Transfer with Importance-Guided Supervision for Fundus Image Enhancement } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16888},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
1.Proposes a unified semi-supervised framework that leverages both unpaired real clinical images and synthetic paired data to achieve state-of-the-art enhancement results. 2.Introduces a knowledge distillation-driven unsupervised quality transfer mechanism that aligns feature representations, bridging the global domain gap without the need for paired training data. 3.Designs a pixel-level, importance-guided supervised learning branch that dynamically adjusts supervision intensity, prioritizing clinically critical regions like vessels and lesions during reconstruction.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
1.Elegant Domain Gap Resolution: The knowledge distillation approach effectively overcomes the reliance on scarce paired clinical data while actively mitigating the structural artifacts commonly associated with global cycle-consistent adversarial networks. 2.Clinically Meaningful Attention: The importance decoder serves as an effective, weakly-supervised attention mechanism. It identifies and preserves diagnostically critical structures such as microaneurysms and hard exudates without requiring expensive, explicit pixel-level annotations.
- 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.Risk of Clinical Hallucination: The strong reliance on uncertainty-guided supervision to restore fine details poses a significant risk of structural hallucination. In regions where the underlying physiological signal is completely obliterated by severe artifacts, the generative model may fabricate realistic but anatomically incorrect vessels or pseudo-lesions (false positives) to satisfy the loss function. This is a critical safety concern for downstream clinical diagnostics. 2.Dependence on Synthetic Degradation: While the unsupervised branch gracefully handles real unpaired data, the supervised branch still relies on a synthetic degradation pipeline to create training pairs. Synthetic degradation models may not encapsulate the full, complex reality of clinical image artifacts. Consequently, the network’s uncertainty estimation is inherently biased towards synthetic patterns. 3.Outdated Baselines: With the exception of Diff-Retinex++ (2025), most of the compared methods were published in or before 2022.The evaluation would be more convincing if it included comparisons with more recent state-of-the-art approaches.
- 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?
Despite valid concerns regarding potential generative hallucinations and the reliance on synthetic degradation, the manuscript is well-written, and the proposed methodology is adequately validated through downstream clinical tasks, demonstrating sufficient practical utility to warrant a weak accept.
- 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 no further question about the rebuttal
Review #2
- Please describe the contribution of the paper
The paper proposes a semi‑supervised dual framework for enhancing low‑quality fundus images, addressing the reliance on paired training data and spatially uniform enhancement. It contains two components: an unsupervised knowledge distillation‑driven quality transfer module, and an importance‑guided supervised learning branch. Experiments demonstrate state‑of‑the‑art enhancement performance and significant improvements in downstream DR grading. Ablation studies verify the effectiveness of proposed components.
- 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 author proposes a novel semi-supervised method that integrates unsupervised quality transfer method and importance-guided supervised enhancement, which mitigate the reliance of paired dataset while enhances the image quality. 2.Experiments show significant performance gains on both image quality and downstream DR grading.
- 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.Image quality is a rather general concept in computer vision. To tackle quality enhancement, it is recommended to clearly define the specify types of quality degradations. Paper mentioned structural artifacts, color distortions, etc., but it is recommended to define it clearly before the method formulation and then indicate the enhancement accordingly in the experiment visualization. 2.The proposed “ importance-guided supervised learning branch” is essentially based on pixel-level reconstruction difficulty. However, it is not reasonable to directly claim” reconstruction difficulty” resigns are important regions for quality enhancements. Such logic gap weakens the motivation of the proposed method, even the experiment results are satisfactory. It is recommended to provided more strict explanation to mitigate this gap.
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission does not mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure reproducibility.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The paper is well-written. The motivation and contributions are clear. The experiments are solid and shows significant improvements. However, there are some major logical gaps needed to be clearly elaborated.
- 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.
Authors explained to my questions and mitigate my concerns, especially the confusion about logic gaps between “importance”, “uncertainty” and “quality”, and their clinical significance. Authors should thoroughly revise the manuscript according to the rebuttal.
Review #3
- Please describe the contribution of the paper
A knowledge distillation-driven quality transfer mechanism aligns generator features with a high-quality domain CNN, bridging the global quality gap without paired data and mitigating artifacts of cycle-consistency-based methods. A pixel-level importance-guided supervised branch incorporates a lightweight importance decoder to adaptively concentrate supervision on clinically critical regions including the optic disc, retinal vessels, and DR lesions.
- 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 addresses a clinically meaningful and practically relevant problem. The motivation is clear, and the distinction between global quality alignment and spatially heterogeneous reconstruction difficulty is sensible. The empirical results on EyeQ are competitive: the method reportedly achieves 27.86 dB PSNR and 0.91 SSIM, outperforming all listed baselines in Table 1.The ablation study also suggests that the proposed importance-guided branch contributes substantially to fine-structure recovery, with a reported +2.62 dB gain over the variant without this branch.
Another positive aspect is that the method does not rely purely on paired real clinical data, which are difficult to obtain in practice. The semi-supervised formulation is therefore reasonable for this application. The paper also provides a reasonably clear high-level description of the training setup, including datasets, backbone choices, image resolution, optimizer, learning rate, batch size, training duration, and loss weights.
- 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 second branch is repeatedly described as “importance-guided” and claimed to prioritize clinically important regions. However, the actual formulation is based on uncertainty-weighted supervision following Kendall and Gal, where higher uncertainty leads to lower loss weight. As a result, the current presentation creates a mismatch between the claimed “importance-guided” mechanism and the implemented uncertainty-based weighting scheme. 2.The paper does not explicitly introduce clinical priors, anatomical annotations, or lesion supervision. Instead, the highlighted regions emerge from learned reconstruction uncertainty. Therefore, it is not fully justified to frame this module as prioritizing clinically important regions in a direct sense. 3.The manuscript introduces an importance map and also uses a per-pixel log-uncertainty term, but the relationship between these quantities is not clearly specified. This makes the core supervision mechanism harder to follow and weakens the technical clarity of the method section. 4.When the shared classifier is first introduced, the paper does not immediately clarify what exactly it classifies. Although later text suggests that the classifier outputs probabilities over disease categories, this should be stated explicitly at the first introduction for better readability and technical transparency. 5.The paper repeatedly emphasizes improvements in downstream DR grading, but the actual grading results are deferred to the supplementary material. Since the supplementary file is unavailable in the review package visible here, it is not possible to assess how strong or convincing the downstream gains really are. This weakens one of the paper’s central claims. 6.It is not sufficiently explained how PSNR/SSIM are computed on EyeQ, given the use of “Good” images as references and “Reject” images as low-quality inputs. The paper would benefit from a more rigorous description of the evaluation setup. 7.The alignment between the high-quality CNN, the low-quality CNN, and the frozen classifier seems to primarily constrain disease-related semantic representations, i.e., it encourages the enhanced image to preserve classification-relevant features. However, this does not necessarily imply that genuine image-quality characteristics have been transferred. In other words, semantic consistency in the classifier space should not be directly equated with quality transfer in the visual domain.
- 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.
(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 gave this paper a Weak Reject mainly because the paper’s central methodological claims are not yet fully convincing. In particular, the second branch is framed as “importance-guided” and clinically aware, but its actual formulation is uncertainty-based, without explicit clinical priors, and the relationship between the importance map and the uncertainty term is not clearly explained. In addition, the distillation objective seems to mainly enforce disease-related semantic consistency through fixed CNNs and a classifier, which does not necessarily establish genuine image-quality transfer. Finally, a key claim of the paper is improved downstream DR grading, but this cannot be properly assessed from the currently accessible materials because the supplementary file is unavailable, and the evaluation setup is not described clearly enough in the main paper.
- 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.
Although the authors provided additional explanations, they failed to address my core concern regarding the fairness of the baseline comparisons. The novelty remains limited. Thus, I maintain my negative rating.
Author Feedback
We sincerely thank all three reviewers and the PC for their constructive feedback. 1)Hallucination Risk(PCQ1;R1Q1-2;R2Q2) a) Structural regularisation by design The perceptual loss LP compares the enhanced image with the low-quality input rather than ground truth, preserving structural consistency via VGG-16 and discouraging nonexistent structures. The total variation loss LTV suppresses hallucinated high-frequency edges. b) Uncertainty-weighted loss suppresses ambiguous regions. The importance decoder assigns high uncertainty (high αi) to regions where the input signal is most ambiguous. Per the Kendall-Gal formulation, high αi suppresses the MSE gradient via 1/exp(αi)→0, concentrating learning on well-defined structures where the input signal is reliable. c) Synthetic degradation coverage The unsupervised branch directly uses real unpaired clinical images, aligning feature distributions between real high- and low-quality domains without synthetic degradation. The supervised branch complements rather than replaces this real-data alignment. 2) Clarifying Importance-Guided and Clinical Relevance(PCQ2;R2Q2;R3Q1–3) The module is mechanistically uncertainty-based, yet its effect is importance-guided. The importance decoder ρ(·) generates a per-pixel uncertainty map α(ŷ), where high αi indicates difficult-to-reconstruct regions. In fundus images, these regions correspond to clinically important structures, such as vessels and lesions, while homogeneous backgrounds yield low αi. Fig. 5 confirms this alignment across all DR severity levels. The loss 1/exp(αi)·LMSE + αi suppresses unreliable regions while focusing learning on reliable structures, and the αi regularisation term prevents the trivial all-uncertain solution. We will revise the framing in the camera-ready version to foreground the uncertainty mechanism and clarify that clinical relevance emerges as an empirical consequence rather than an explicit prior. 3) Visual Quality Transfer(PCQ3;R3Q7) LKD is not the sole quality-transfer mechanism. Visual quality is further enforced by: LRP, capturing global texture and style through Gram-matrix channel correlations; LP, preserving structural details via VGG-16 perceptual features; and LTV, suppressing noise. Moreover, the shared ResNet-50 backbone encodes domain-specific visual characteristics including illumination and colour calibration, so aligning p(ft(x)) with p(fs(x̃)) implicitly maps low-quality images into the high-quality visual domain. This is supported by ablation: removing LKD introduces a pronounced warm colour cast, a visual artefact unrelated to semantic labels, confirming that LKD contributes to visual, not merely semantic, alignment. 4)Downstream Task and Evaluation Protocol (PCQ4;R3Q5&6) A DenseNet-121 classifier was trained for four-level DR grading on enhanced Messidor images. All methods enhanced the same images before classification. Our method achieved ACC 0.737 / QWK 0.723, improving over the unenhanced baseline (ACC 0.681 / QWK 0.657) by +5.6% ACC and +6.6% QWK, outperforming the closest competitor I-SECRET (ACC 0.735 / QWK 0.718). Regarding the EyeQ protocol: following Shen et al. [18], “Good” images served as supervised references and PSNR/SSIM ground truth; “Reject” images served as real low-quality test inputs; “Usable” images were excluded. PSNR/SSIM were computed between the enhanced output and the corresponding “Good” image in each synthetic pair. 5)Additional Points R1Q3:Diff-Retinex++ was the latest SOTA enhancement method available at submission, and our method surpasses it by +1.83 dB PSNR. More recent methods will be added in the camera-ready version. R2Q1:Degradation Types. The addressed degradations include illumination imbalance, optical blur, and low SNR in dark peripheral regions, representing the most common clinical fundus image failure modes. R3Q4:Shared Classifier Definition. p(·) denotes a shared DR classifier predicting grades 0–4, with frozen parameters during generator training.
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 main concerns are with methodological clarity and logic. Several points should be addressed in the rebuttal, but not limited to:
1.Discuss the risk of generating anatomically incorrect structures / false lesions under severe artifacts, and the potential bias introduced by reliance on synthetic degradation modeling. 2.Clarify the claim of “importance-guided” and “clinical priors”. 3.Semantic consistency in the classifier space should not be directly equated with quality transfer in the visual domain. This needs to be justified. 4.The claimed downstream DR grading improvement is not sufficiently supported in the main paper, which weakens this central claim. The authors should clarify the evaluation and provide stronger evidence or moderate the claim accordingly.
- 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 rebuttal addressed many of the main concerns and led to an overall positive post-rebuttal assessment. In particular, the authors clarified the relationship between uncertainty, importance, and clinical relevance, moderated the interpretation of the guidance mechanism, and clarified the downstream grading evaluation protocol. One reviewer still raises a valid concern regarding the fairness of the baseline comparisons and the overall degree of methodological novelty. Nevertheless, considering the overall reviewer feedback after rebuttal, the paper appears slightly above the acceptance threshold.
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.
After considering the rebuttal and all reviewers’ comments, the overall assessment remains positive. The rebuttal provides useful clarifications, and the paper’s methodological design and experimental results support acceptance.
