arxiv: 2604.03806 · v1 · submitted 2026-04-04 · 💻 cs.CV

Recognition: no theorem link

Bridging Restoration and Diagnosis: A Comprehensive Benchmark for Retinal Fundus Enhancement

Xuanzhao Dong , Wenhui Zhu , Xiwen Chen , Hao Wang , Xin Li , Yujian Xiong , Jiajun Cheng , Zhipeng Wang

show 3 more authors

Shao Tang Oana Dumitrascu Yalin Wang

Authors on Pith no claims yet

Pith reviewed 2026-05-13 17:11 UTC · model grok-4.3

classification 💻 cs.CV

keywords fundus image enhancementretinal benchmarkclinical evaluationvessel segmentationdiabetic retinopathy gradingexpert assessmentgenerative modelsimage restoration

0 comments

The pith

EyeBench-V2 evaluates fundus enhancement models by clinical task performance and expert review rather than pixel metrics.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper introduces EyeBench-V2 to address gaps in how generative models for retinal fundus image enhancement are judged. Standard metrics like PSNR and SSIM miss clinically relevant aspects such as lesion preservation and consistent vessel morphology. The benchmark adds downstream evaluations on vessel segmentation, diabetic retinopathy grading, lesion segmentation, and generalization to unseen noise, paired with a structured expert assessment protocol on a curated dataset that checks for alterations in lesion structure, background color shifts, and artificial features. This setup supplies actionable comparisons for both paired and unpaired enhancement methods to steer development toward clinically useful outputs.

Core claim

EyeBench-V2 bridges restoration and diagnosis by supplying a unified benchmark that measures enhancement models through multi-dimensional clinical alignment, including performance on vessel segmentation, DR grading, and lesion segmentation plus expert manual review of lesion alterations, color shifts, and introduced artifacts on a dataset supporting fair paired and unpaired comparisons.

What carries the argument

EyeBench-V2 benchmark with its downstream clinical tasks and expert-guided manual assessment protocol that enables unified evaluation of enhancement methods.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The benchmark could drive training objectives that directly optimize for segmentation and grading accuracy rather than image similarity alone.
Similar task-oriented benchmarks might be applied to enhancement in other medical imaging areas to link restoration more tightly to diagnostic outcomes.
Current generative models may need architectural changes to avoid introducing artifacts that the expert protocol flags as clinically harmful.

Load-bearing premise

The chosen downstream tasks and expert assessment protocol accurately reflect real clinical diagnostic utility and the dataset represents typical real-world fundus images and noise.

What would settle it

A controlled reader study showing that models ranked highest by EyeBench-V2 produce no measurable gain in ophthalmologist diagnostic accuracy on enhanced images compared with originals.

Figures

Figures reproduced from arXiv: 2604.03806 by Hao Wang, Jiajun Cheng, Oana Dumitrascu, Shao Tang, Wenhui Zhu, Xin Li, Xiwen Chen, Xuanzhao Dong, Yalin Wang, Yujian Xiong, Zhipeng Wang.

**Figure 1.** Figure 1: Overview of EyeBench-v2. We present EyeBench-v2, a systematic and comprehensive benchmark designed to evaluate retinal image enhancement models. The evaluation pipeline encompasses both Full-Reference and No-Reference assessments, enabling a robust multi-dimensional analysis of enhancement quality. For each evaluation aspect, we construct a distribution-aligned dataset to ensure fair, reproducible, and cli… view at source ↗

**Figure 2.** Figure 2: T-SNE [34] visualizations of the latent representation features extracted from the RETfound (A) [41] and Ret-Clip (B) [8] image encoder in no-reference evaluation. Here, blue points illustrate synthetic high-quality image yˆ1 features while green points show true high-quality image y2 features. Closer proximity of the distributions indicates improved denoising performance of the unpaired method. More detai… view at source ↗

**Figure 3.** Figure 3: Validation of Expert Clinic Preference Alignment via Spearman’s correlation coefficient ( [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗

**Figure 4.** Figure 4: Illustration of the limitations of SDE-based methods. [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗

**Figure 5.** Figure 5: Overview of EyeQ [11] dataset. (A) highlights attribute distributions (i.e., brightness, contrast, sharpness) and diabetic retinopathy (DR) grades across quality categories (i.e., good, usable, and reject). (B) illustrates histograms for the training (i.e., part A and part B), testing, and validation datasets used in Full-Reference evaluations after resampling, with the workflow of degradation algorithms o… view at source ↗

**Figure 6.** Figure 6: An illustrative medical expert clinical preference eval [PITH_FULL_IMAGE:figures/full_fig_p013_6.png] view at source ↗

**Figure 7.** Figure 7: Illustration of the Denoising Evaluation on the EyeQ dataset. The first and second columns show the high- and low-quality image [PITH_FULL_IMAGE:figures/full_fig_p014_7.png] view at source ↗

**Figure 8.** Figure 8: Illustration of the Denoising Generalization Evaluation on the DRIVE and IDRID datasets. The first and second columns show [PITH_FULL_IMAGE:figures/full_fig_p014_8.png] view at source ↗

**Figure 9.** Figure 9: Illustration of Vessel and Lesion (EX and HE) Segmentation Experiments. The first column shows the reference segmentation [PITH_FULL_IMAGE:figures/full_fig_p015_9.png] view at source ↗

**Figure 10.** Figure 10: Illustration of the denoising results in the No-Reference Quality Assessment Experiments. The first column shows the input [PITH_FULL_IMAGE:figures/full_fig_p015_10.png] view at source ↗

read the original abstract

Over the past decade, generative models have demonstrated success in enhancing fundus images. However, the evaluation of these models remains a challenge. A benchmark for fundus image enhancement is needed for three main reasons:(1) Conventional denoising metrics such as PSNR and SSIM fail to capture clinically relevant features, such as lesion preservation and vessel morphology consistency, limiting their applicability in real-world settings; (2) There is a lack of unified evaluation protocols that address both paired and unpaired enhancement methods, particularly those guided by clinical expertise; and (3) An evaluation framework should provide actionable insights to guide future advancements in clinically aligned enhancement models. To address these gaps, we introduce EyeBench-V2, a benchmark designed to bridge the gap between enhancement model performance and clinical utility. Our work offers three key contributions:(1) Multi-dimensional clinical-alignment through downstream evaluations: Beyond standard enhancement metrics, we assess performance across clinically meaningful tasks including vessel segmentation, diabetic retinopathy (DR) grading, generalization to unseen noise patterns, and lesion segmentation. (2) Expert-guided evaluation design: We curate a novel dataset enabling fair comparisons between paired and unpaired enhancement methods, accompanied by a structured manual assessment protocol by medical experts, which evaluates clinically critical aspects such as lesion structure alterations, background color shifts, and the introduction of artificial structures. (3) Actionable insights: Our benchmark provides a rigorous, task-oriented analysis of existing generative models, equipping clinical researchers with the evidence needed to make informed decisions, while also identifying limitations in current methods to inform the design of next-generation enhancement models.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

EyeBench-V2 sets up a clinically oriented benchmark for fundus enhancement with downstream tasks and expert review, but the description supplies no results or validation data to show the proxies actually track diagnostic value.

read the letter

The main thing to know is that this paper introduces EyeBench-V2, a benchmark that evaluates fundus image enhancement models on vessel segmentation, DR grading, lesion segmentation, and a structured expert protocol covering lesion structure, color shifts, and artificial artifacts, while supporting both paired and unpaired methods on a curated dataset. The goal is to move past PSNR and SSIM toward metrics that matter for clinical use. That framing is reasonable and directly tackles a known shortcoming in medical image work. The setup for fair comparisons across method types and the inclusion of generalization to unseen noise are clear improvements over generic benchmarks. The paper also does a straightforward job listing why current metrics fall short on lesion preservation and vessel morphology. The soft spot is the missing evidence. The text describes the framework and protocol but reports no quantitative results, no dataset statistics, no inter-rater numbers, and no correlation checks between the chosen tasks or expert scores and actual diagnostic accuracy. Without that link, the claim that the benchmark bridges enhancement performance to clinical utility stays untested. The stress-test concern lands: if the proxies only loosely relate to real decisions across cameras or patient groups, the rankings and insights lose force. This is for researchers working on generative models for retinal images who need evaluation tools that reflect ophthalmology priorities. A reader in medical AI would pick up the protocol details and the analysis of existing model limits. It deserves a serious referee because the gap it targets is real and the design is specific enough to review, though the authors will need to show the validation data in revisions.

Referee Report

3 major / 1 minor

Summary. The manuscript introduces EyeBench-V2, a benchmark for retinal fundus image enhancement. It argues that standard metrics such as PSNR and SSIM fail to capture clinically relevant features like lesion preservation and vessel morphology. The benchmark evaluates generative models through downstream clinical tasks including vessel segmentation, diabetic retinopathy grading, lesion segmentation, and generalization to unseen noise patterns, supplemented by expert-guided manual assessment on a curated dataset supporting paired and unpaired method comparisons. The goal is to deliver actionable insights that align enhancement performance with diagnostic utility and guide future model development.

Significance. If the chosen downstream tasks and expert protocol are validated as reliable proxies, EyeBench-V2 could establish a much-needed standardized evaluation framework in ophthalmic image enhancement. This would help the community move beyond generic image-quality metrics toward assessments that better reflect real diagnostic value, potentially improving model selection for clinical deployment and highlighting specific weaknesses in current generative approaches.

major comments (3)

[Abstract] Abstract, point (1): The claim that downstream tasks (vessel segmentation, DR grading, lesion segmentation) bridge enhancement performance to clinical utility is load-bearing but unsupported. No correlation analysis, inter-rater reliability statistics, or evidence linking task improvements to actual diagnostic accuracy is provided, leaving the proxy assumption untested.
[Contributions] Contributions (2): The expert-guided evaluation design is described at a high level but lacks concrete protocol details such as scoring rubrics for lesion structure alterations, color shifts, and artificial structures, expert selection criteria, or quantitative agreement measures. Without these, reproducibility and clinical fidelity cannot be assessed.
[Contributions] Contributions (3) and dataset description: The assertion of 'actionable insights' and fair paired/unpaired comparisons depends on the curated dataset representing real-world noise patterns across camera types and populations. No statistics on dataset size, diversity, or noise modeling are supplied, weakening the generalization claims.

minor comments (1)

[Abstract] The abstract mentions 'generalization to unseen noise patterns' without defining the patterns or the protocol used to introduce them; adding one sentence of clarification would improve readability.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the thoughtful and constructive review of our manuscript on EyeBench-V2. The comments highlight important areas for strengthening the presentation of our clinical-alignment claims, evaluation protocols, and dataset details. We address each major comment below and outline the specific revisions planned for the next version of the paper.

read point-by-point responses

Referee: [Abstract] Abstract, point (1): The claim that downstream tasks (vessel segmentation, DR grading, lesion segmentation) bridge enhancement performance to clinical utility is load-bearing but unsupported. No correlation analysis, inter-rater reliability statistics, or evidence linking task improvements to actual diagnostic accuracy is provided, leaving the proxy assumption untested.

Authors: We agree that an explicit correlation analysis between enhancement outputs and downstream diagnostic accuracy would provide stronger support for the proxy assumption. The selected tasks (vessel segmentation, DR grading, lesion segmentation) were chosen because they are established clinical endpoints in ophthalmology literature, directly tied to diagnostic decisions. In the revision we will add a dedicated paragraph in the abstract and methods sections justifying these proxies with supporting references, report inter-rater reliability statistics for the expert assessments, and include a supplementary correlation table (e.g., Spearman rank between perceptual scores and task metrics) computed on the existing evaluation data. This will make the bridging claim more robust without requiring new experiments. revision: partial
Referee: [Contributions] Contributions (2): The expert-guided evaluation design is described at a high level but lacks concrete protocol details such as scoring rubrics for lesion structure alterations, color shifts, and artificial structures, expert selection criteria, or quantitative agreement measures. Without these, reproducibility and clinical fidelity cannot be assessed.

Authors: We appreciate this observation. The original description was kept concise, but we recognize that full reproducibility requires the complete protocol. In the revised manuscript we will expand the expert evaluation subsection to include: (i) the full 5-point scoring rubrics for lesion structure, color fidelity, and artifact introduction; (ii) expert selection criteria (board-certified ophthalmologists with at least five years of fundus-image experience); and (iii) quantitative agreement metrics (Cohen’s kappa and percentage agreement) computed across the three experts. These additions will be placed in the main text with the rubrics also provided as supplementary material. revision: yes
Referee: [Contributions] Contributions (3) and dataset description: The assertion of 'actionable insights' and fair paired/unpaired comparisons depends on the curated dataset representing real-world noise patterns across camera types and populations. No statistics on dataset size, diversity, or noise modeling are supplied, weakening the generalization claims.

Authors: We acknowledge the need for explicit dataset statistics. The revised version will add a new table and accompanying text detailing: total image count, breakdown by camera manufacturer (Topcon, Zeiss, Canon, etc.), patient demographics (age range, ethnicity distribution where available), and the noise-modeling procedure (synthetic noise calibrated to real acquisition artifacts observed across the source cameras). These statistics will directly support the claims of real-world representativeness and fair paired/unpaired comparisons. revision: yes

Circularity Check

0 steps flagged

No circularity: benchmark proposal uses external proxies without self-referential derivation

full rationale

The paper introduces EyeBench-V2 as an evaluation framework for fundus enhancement models, relying on downstream tasks (vessel segmentation, DR grading, lesion segmentation) and an expert assessment protocol. No mathematical derivations, equations, or parameter fittings exist that could reduce to self-definition or fitted inputs called predictions. The central premise—that these tasks and protocols bridge to clinical utility—is presented as an assumption supported by curation choices, not derived internally or justified via self-citation chains. No uniqueness theorems, ansatzes smuggled through citations, or renamings of known results appear. The work is self-contained as a benchmark definition whose validity hinges on external clinical correlation, not circular construction.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 0 invented entities

The benchmark rests on domain assumptions about clinical relevance rather than new mathematical constructs or fitted parameters.

axioms (2)

domain assumption Downstream tasks such as vessel segmentation and DR grading serve as valid indicators of clinical utility for enhanced fundus images.
Invoked when defining the multi-dimensional clinical-alignment evaluations.
domain assumption Structured expert review can reliably detect clinically critical alterations such as lesion structure changes or artificial structures.
Central to the expert-guided evaluation design described in contribution (2).

pith-pipeline@v0.9.0 · 5619 in / 1222 out tokens · 43353 ms · 2026-05-13T17:11:25.421159+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

47 extracted references · 47 canonical work pages · 1 internal anchor

[1]

Wasserstein generative adversarial networks

Martin Arjovsky, Soumith Chintala, and L ´eon Bottou. Wasserstein generative adversarial networks. InInterna- tional conference on machine learning, pages 214–223. PMLR, 2017. 3

work page 2017
[2]

On the mathematical properties of the structural similarity index

Dominique Brunet, Edward R Vrscay, and Zhou Wang. On the mathematical properties of the structural similarity index. IEEE Transactions on Image Processing, 21(4):1488–1499,

work page
[3]

I-secret: Importance-guided fundus image enhance- ment via semi-supervised contrastive constraining

Pujin Cheng, Li Lin, Yijin Huang, Junyan Lyu, and Xiaoying Tang. I-secret: Importance-guided fundus image enhance- ment via semi-supervised contrastive constraining. InMed- ical Image Computing and Computer Assisted Intervention– MICCAI 2021: 24th International Conference, Strasbourg, France, September 27–October 1, 2021, Proceedings, Part VIII 24, pages ...

work page 2021
[4]

Improving rep- resentation of high-frequency components for medical foun- dation models.arXiv preprint arXiv:2407.14651, 2024

Yuetan Chu, Yilan Zhang, Zhongyi Han, Changchun Yang, Longxi Zhou, Gongning Luo, and Xin Gao. Improving rep- resentation of high-frequency components for medical foun- dation models.arXiv preprint arXiv:2407.14651, 2024. 8

work page arXiv 2024
[5]

Zhuo Deng, Yuanhao Cai, Lu Chen, Zheng Gong, Qiqi Bao, Xue Yao, Dong Fang, Wenming Yang, Shaochong Zhang, and Lan Ma. Rformer: Transformer-based generative ad- versarial network for real fundus image restoration on a new clinical benchmark.IEEE Journal of Biomedical and Health Informatics, 26(9):4645–4655, 2022. 3, 4, 5, 12

work page 2022
[6]

Cunsb-rfie: Context-aware unpaired neural schr”{o}dinger bridge in retinal fundus im- age enhancement.arXiv preprint arXiv:2409.10966, 2024

Xuanzhao Dong, Vamsi Krishna Vasa, Wenhui Zhu, Peijie Qiu, Xiwen Chen, Yi Su, Yujian Xiong, Zhangsihao Yang, Yanxi Chen, and Yalin Wang. Cunsb-rfie: Context-aware unpaired neural schr”{o}dinger bridge in retinal fundus im- age enhancement.arXiv preprint arXiv:2409.10966, 2024. 2, 3, 4, 5, 7, 8, 13

work page arXiv 2024
[7]

Tpot: Topology pre- serving optimal transport in retinal fundus image enhance- ment.arXiv preprint arXiv:2411.01403, 2024

Xuanzhao Dong, Wenhui Zhu, Xin Li, Guoxin Sun, Yi Su, Oana M Dumitrascu, and Yalin Wang. Tpot: Topology pre- serving optimal transport in retinal fundus image enhance- ment.arXiv preprint arXiv:2411.01403, 2024. 3, 4, 5, 13

work page arXiv 2024
[8]

Ret-clip: A retinal im- age foundation model pre-trained with clinical diagnostic re- ports.arXiv preprint arXiv:2405.14137, 2024

Jiawei Du, Jia Guo, Weihang Zhang, Shengzhu Yang, Han- ruo Liu, Huiqi Li, and Ningli Wang. Ret-clip: A retinal im- age foundation model pre-trained with clinical diagnostic re- ports.arXiv preprint arXiv:2405.14137, 2024. 5, 6, 7, 14

work page arXiv 2024
[9]

Di- abetic retinopathy detection.https : / / kaggle

Emma Dugas, Jared, Jorge, and Will Cukierski. Di- abetic retinopathy detection.https : / / kaggle . com / competitions / diabetic - retinopathy - detection, 2015. Kaggle. 11

work page 2015
[10]

Automated retinal imaging anal- ysis for alzheimers disease screening

Oana M Dumitrascu, Wenhui Zhu, Peijie Qiu, Keshav Nan- dakumar, and Yalin Wang. Automated retinal imaging anal- ysis for alzheimers disease screening. InIEEE International Symposium on Biomedical Imaging: From Nano to Macro (ISBI), 2022. 2

work page 2022
[11]

Evaluation of reti- nal image quality assessment networks in different color- spaces

Huazhu Fu, Boyang Wang, Jianbing Shen, Shanshan Cui, Yanwu Xu, Jiang Liu, and Ling Shao. Evaluation of reti- nal image quality assessment networks in different color- spaces. InMedical Image Computing and Computer Assisted Intervention–MICCAI 2019: 22nd International Conference, Shenzhen, China, October 13–17, 2019, Proceedings, Part I 22, pages 48–56. Sp...

work page 2019
[12]

Generative adversarial networks.Commu- nications of the ACM, 63(11):139–144, 2020

Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks.Commu- nications of the ACM, 63(11):139–144, 2020. 3

work page 2020
[13]

Improved training of wasserstein gans.Advances in neural information processing systems, 30, 2017

Ishaan Gulrajani, Faruk Ahmed, Martin Arjovsky, Vincent Dumoulin, and Aaron C Courville. Improved training of wasserstein gans.Advances in neural information processing systems, 30, 2017. 3, 4, 5, 12

work page 2017
[14]

Denoising dif- fusion probabilistic models.Advances in neural information processing systems, 33:6840–6851, 2020

Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising dif- fusion probabilistic models.Advances in neural information processing systems, 33:6840–6851, 2020. 3

work page 2020
[15]

Image-to-image translation with conditional adver- sarial networks

Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros. Image-to-image translation with conditional adver- sarial networks. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 1125–1134,

work page
[16]

Unpaired image-to-image translation via neural schr\” odinger bridge.arXiv preprint arXiv:2305.15086, 2023

Beomsu Kim, Gihyun Kwon, Kwanyoung Kim, and Jong Chul Ye. Unpaired image-to-image translation via neu- ral schr\” odinger bridge.arXiv preprint arXiv:2305.15086,

work page arXiv
[17]

Structure-consistent restoration network for cataract fundus image enhancement

Heng Li, Haofeng Liu, Huazhu Fu, Hai Shu, Yitian Zhao, Xiaoling Luo, Yan Hu, and Jiang Liu. Structure-consistent restoration network for cataract fundus image enhancement. InInternational Conference on Medical Image Comput- ing and Computer-Assisted Intervention, pages 487–496. Springer, 2022. 3, 4, 5, 11

work page 2022
[18]

A generic fundus image enhancement network boosted by frequency self-supervised representation learning.Medical Image Analysis, 90:102945,

Heng Li, Haofeng Liu, Huazhu Fu, Yanwu Xu, Hai Shu, Ke Niu, Yan Hu, and Jiang Liu. A generic fundus image enhancement network boosted by frequency self-supervised representation learning.Medical Image Analysis, 90:102945,

work page
[19]

Degradation-invariant enhance- ment of fundus images via pyramid constraint network

Haofeng Liu, Heng Li, Huazhu Fu, Ruoxiu Xiao, Yunshu Gao, Yan Hu, and Jiang Liu. Degradation-invariant enhance- ment of fundus images via pyramid constraint network. In Medical Image Computing and Computer Assisted Interven- tion – MICCAI 2022, pages 507–516, Cham, 2022. Springer Nature Switzerland. 3, 4, 5, 11

work page 2022
[20]

Least squares genera- tive adversarial networks

Xudong Mao, Qing Li, Haoran Xie, Raymond YK Lau, Zhen Wang, and Stephen Paul Smolley. Least squares genera- tive adversarial networks. InProceedings of the IEEE inter- national conference on computer vision, pages 2794–2802,

work page
[21]

The contextual loss for image transformation with non-aligned data, 2018

Roey Mechrez, Itamar Talmi, and Lihi Zelnik-Manor. The contextual loss for image transformation with non-aligned data, 2018. 3, 13

work page 2018
[22]

Contrastive learning for unpaired image-to-image translation

Taesung Park, Alexei A Efros, Richard Zhang, and Jun- Yan Zhu. Contrastive learning for unpaired image-to-image translation. InComputer Vision–ECCV 2020: 16th Euro- pean Conference, Glasgow, UK, August 23–28, 2020, Pro- ceedings, Part IX 16, pages 319–345. Springer, 2020. 13

work page 2020
[23]

Idrid: A database for diabetic retinopathy screening research.Data, 3(3), 2018

Prasanna Porwal and et al. Idrid: A database for diabetic retinopathy screening research.Data, 3(3), 2018. 4, 15

work page 2018
[24]

A competition for the diagnosis of my- opic maculopathy by artificial intelligence algorithms.JAMA ophthalmology, 2024

Bo Qian, Bin Sheng, Hao Chen, Xiangning Wang, Tingyao Li, Yixiao Jin, Zhouyu Guan, Zehua Jiang, Yilan Wu, Jinyuan Wang, et al. A competition for the diagnosis of my- opic maculopathy by artificial intelligence algorithms.JAMA ophthalmology, 2024. 2

work page 2024
[25]

Medical image segmentation based on frequency domain decomposition svd linear attention.Scientific Re- ports, 15(1):2833, 2025

Liu Qiong, Li Chaofan, Teng Jinnan, Chen Liping, and Song Jianxiang. Medical image segmentation based on frequency domain decomposition svd linear attention.Scientific Re- ports, 15(1):2833, 2025. 8

work page 2025
[26]

High-resolution image synthesis with latent diffusion models

Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bj ¨orn Ommer. High-resolution image synthesis with latent diffusion models. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684–10695, 2022. 7

work page 2022
[27]

U-net: Convolutional networks for biomedical image segmentation,

Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation,

work page
[28]

Can push-forward generative models fit multimodal distributions?Advances in Neural Information Processing Systems, 35:10766–10779, 2022

Antoine Salmona, Valentin De Bortoli, Julie Delon, and Agnes Desolneux. Can push-forward generative models fit multimodal distributions?Advances in Neural Information Processing Systems, 35:10766–10779, 2022. 3

work page 2022
[29]

Mod- eling and enhancing low-quality retinal fundus images.IEEE transactions on medical imaging, 40(3):996–1006, 2020

Ziyi Shen, Huazhu Fu, Jianbing Shen, and Ling Shao. Mod- eling and enhancing low-quality retinal fundus images.IEEE transactions on medical imaging, 40(3):996–1006, 2020. 2, 3, 4, 5, 11

work page 2020
[30]

Freeu: Free lunch in diffusion u-net

Chenyang Si, Ziqi Huang, Yuming Jiang, and Ziwei Liu. Freeu: Free lunch in diffusion u-net. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4733–4743, 2024. 8

work page 2024
[31]

Denoising Diffusion Implicit Models

Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models.arXiv preprint arXiv:2010.02502, 2020. 3

work page internal anchor Pith review Pith/arXiv arXiv 2010
[32]

Generative modeling by esti- mating gradients of the data distribution.Advances in neural information processing systems, 32, 2019

Yang Song and Stefano Ermon. Generative modeling by esti- mating gradients of the data distribution.Advances in neural information processing systems, 32, 2019. 3

work page 2019
[33]

Staal and et al

J. Staal and et al. Ridge-based vessel segmentation in color images of the retina.IEEE Trans Med Imaging, 23(4):501– 509, 2004. 4, 15

work page 2004
[34]

Visualizing data using t-sne.Journal of machine learning research, 9 (11), 2008

Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne.Journal of machine learning research, 9 (11), 2008. 6, 7

work page 2008
[35]

Context-aware opti- mal transport learning for retinal fundus image enhancement

Vamsi Krishna Vasa, Peijie Qiu, Wenhui Zhu, Yujian Xiong, Oana Dumitrascu, and Yalin Wang. Context-aware opti- mal transport learning for retinal fundus image enhancement. arXiv preprint arXiv:2409.07862, 2024. 2, 3, 4, 5, 12, 13

work page arXiv 2024
[36]

Springer, 2009

C ´edric Villani et al.Optimal transport: old and new. Springer, 2009. 3

work page 2009
[37]

Rbad: A dataset and benchmark for retinal bifurcation angle detection

Hao Wang, Wenhui Zhu, Jiayou Qin, Xin Li, Oana Dumi- trascu, Xiwen Chen, Peijie Qiu, Abolfazl Razi, and Yalin Wang. Rbad: A dataset and benchmark for retinal bifurcation angle detection. InIEEE-EMBS International Conference on Biomedical and Health Informatics. 2

work page
[38]

Optimal transport for unsupervised denoising learning.IEEE Trans- actions on Pattern Analysis and Machine Intelligence, 45(2): 2104–2118, 2022

Wei Wang, Fei Wen, Zeyu Yan, and Peilin Liu. Optimal transport for unsupervised denoising learning.IEEE Trans- actions on Pattern Analysis and Machine Intelligence, 45(2): 2104–2118, 2022. 2, 3, 4, 5, 12

work page 2022
[39]

Mul- tiscale structural similarity for image quality assessment

Zhou Wang, Eero P Simoncelli, and Alan C Bovik. Mul- tiscale structural similarity for image quality assessment. In The Thrity-Seventh Asilomar Conference on Signals, Systems & Computers, 2003, pages 1398–1402. Ieee, 2003. 3

work page 2003
[40]

Study group learning: Improving retinal vessel segmentation trained with noisy labels

Yuqian Zhou, Hanchao Yu, and Humphrey Shi. Study group learning: Improving retinal vessel segmentation trained with noisy labels. InMedical Image Computing and Computer Assisted Intervention–MICCAI 2021: 24th International Conference, Strasbourg, France, September 27–October 1, 2021, Proceedings, Part I 24, pages 57–67. Springer, 2021. 13

work page 2021
[41]

A foundation model for generalizable disease detection from retinal images.Nature, 622(7981):156–163,

Yukun Zhou, Mark A Chia, Siegfried K Wagner, Murat S Ayhan, Dominic J Williamson, Robbert R Struyven, Tim- ing Liu, Moucheng Xu, Mateo G Lozano, Peter Woodward- Court, et al. A foundation model for generalizable disease detection from retinal images.Nature, 622(7981):156–163,

work page
[42]

Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A. Efros. Unpaired Image-to-Image Translation Using Cycle- Consistent Adversarial Networks.CVPR, pages 2242–2251,

work page
[43]

Be- yond mobilenet: An improved mobilenet for retinal diseases

Wenhui Zhu, Peijie Qiu, Xiwen Chen, Huayu Li, Hao Wang, Natasha Lepore, Oana M Dumitrascu, and Yalin Wang. Be- yond mobilenet: An improved mobilenet for retinal diseases. InInternational Conference on Medical Image Computing and Computer-Assisted Intervention, pages 56–65. Springer,

work page
[44]

Otre: Where optimal trans- port guided unpaired image-to-image translation meets reg- ularization by enhancing

Wenhui Zhu, Peijie Qiu, Oana M Dumitrascu, Jacob M Sobczak, Mohammad Farazi, Zhangsihao Yang, Keshav Nandakumar, and Yalin Wang. Otre: Where optimal trans- port guided unpaired image-to-image translation meets reg- ularization by enhancing. InInternational Conference on In- formation Processing in Medical Imaging, pages 415–427. Springer, 2023. 2, 3, 12

work page 2023
[45]

Optimal transport guided unsupervised learning for enhancing low- quality retinal images.Proc IEEE Int Symp Biomed Imaging,

Wenhui Zhu, Peijie Qiu, Mohammad Farazi, Keshav Nan- dakumar, Oana M Dumitrascu, and Yalin Wang. Optimal transport guided unsupervised learning for enhancing low- quality retinal images.Proc IEEE Int Symp Biomed Imaging,

work page
[46]

Self-supervised equivariant regu- larization reconciles multiple-instance learning: Joint refer- able diabetic retinopathy classification and lesion segmenta- tion

Wenhui Zhu, Peijie Qiu, Natasha Lepore, Oana M Dumi- trascu, and Yalin Wang. Self-supervised equivariant regu- larization reconciles multiple-instance learning: Joint refer- able diabetic retinopathy classification and lesion segmenta- tion. In18th International Symposium on Medical Informa- tion Processing and Analysis, pages 100–107. SPIE, 2023. 2

work page 2023
[47]

Dumitrascu, and Yalin Wang

Wenhui Zhu, Peijie Qiu, Xiwen Chen, Xin Li, Natasha Lep- ore, Oana M. Dumitrascu, and Yalin Wang. nnmobilenet: Rethinking cnn for retinopathy research. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, pages 2285–2294, 2024. 2, 4, 14 Supplementary Materials - Bridging Restoration and Diagnosis: A Compre-...

work page 2024