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

cs.CVApr 4, 2026

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

Authors: Xuanzhao Dong, Wenhui Zhu, Xiwen Chen, Hao Wang, Xin Li, Yujian Xiong, Jiajun Cheng, Zhipeng Wang, +3 more

Organizations: Arizona State University · Clemson University · LinkedIn Corporation · Mayo Clinic

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.

Figures & tables

	Method	EyeQ		IDRID		DRIVE
	Method	SSIM $\uparrow$	PSNR $\uparrow$	SSIM $\uparrow$	PSNR $\uparrow$	SSIM $\uparrow$	PSNR $\uparrow$
Paired Methods	SCR-Net [ 17 ]	0.9606	29.698	0.6425	18.920	0.6824	23.280
	Cofe-Net [ 29 ]	0.9408	24.907	0.7397	20.058	0.6671	21.774
	PCE-Net [ 19 ]	0.9487	29.895	0.7764	23.201	0.6704	24.041
	GFE-Net [ 18 ]	0.9554	29.719	0.7935	25.012	0.6793	23.786
	RFormer [ 5 ]	0.9260	27.163	0.5963	18.433	0.6311	22.172

Table 1: Performance comparison of denoising evaluation in Full-Reference quality assessment experiments. The best performance in each column is highlighted in bold, with the second-best underlined. Visualization results refer to the Appendix D .

Method	Vessel Segmentation				EX			HE
Method	AUC $\uparrow$	PR $\uparrow$	F1 Score $\uparrow$	SP $\uparrow$	AUC	PR	F1 Score	AUC	PR	F1 Score
SCR-Net [ 17 ]	0.9227	0.7783	0.7000	0.9787	0.9683	0.6041	0.5556	0.9377	0.3213	0.3725
cofe-Net [ 29 ]	0.9188	0.7698	0.6895	0.9801	0.9623	0.5620	0.5349	0.9302	0.3152	0.3281
PCE-Net [ 19 ]	0.9146	0.7616	0.6790	0.9814	0.9667	0.5876	0.5066	0.9545	0.3639	0.3736
GFE-Net [ 18 ]	0.9175	0.7669	0.6832	0.9814	0.9560	0.5548	0.5380	0.9577	0.4113	0.3751
RFormer [ 5 ]	0.8990	0.7239	0.6374	0.9806	0.9626	0.5593	0.4692	0.9207	0.2677	0.3136

Table 2: Performance comparison of vessel and lesion (EX and HE) segmentation in Full-Reference quality assessment experiments. The best performance in each column is highlighted in bold, with the second-best underlined. For visualization results, refer to the Appendix D .

Method	DR grading				Representation Feature		Experts Protocol Evaluation
Method	ACC $\uparrow$	Kappa $\uparrow$	F1 Score $\uparrow$	AUC $\uparrow$	FID-Retfound [ 40 ] $\downarrow$	FID-Clip [ 8 ] $\downarrow$	LPR $\uparrow$	BPR $\uparrow$	SPR $\uparrow$
CycleGAN [ 42 ]	0.7588	0.6006	0.7180	0.9251	23.778	11.530	0.7707	0.8153	0.8726
WGAN [ 13 ]	0.6446	0.3123	0.6156	0.8874	74.885	33.076	0.4076	0.4204	0.6561
OTTGAN [ 38 ]	0.7440	0.5688	0.7037	0.9247	51.201	20.505	0.4586	0.7580	0.5860
OTEGAN [ 46 ]	0.7539	0.6433	0.7228	0.9326	28.987	11.114	0.8280	0.8981	0.6178
Context-aware OT [ 35 ]	0.7301	0.3811	0.6662	0.9112	61.429	34.456	0.3566	0.3121	0.5159

Table 3: Performance comparison with unpaired baselines in No-Reference quality assessment task. The best performance in each column is highlighted in bold, and the second-best is underlined. Visualization results refer to Appendix D .

Figure 2: T-SNE [ 34 ] visualizations of the latent representation features extracted from the RETfound (A) [ 40 ] and Ret-Clip (B) [ 8 ] image encoder in no-reference evaluation. Here, blue points illustrate synthetic high-quality image $\hat{\mathbf{y}}_{1}$ features while green points show true high-quality image $\mathbf{y}_{2}$ features. Closer proximity of the distributions indicates improved denoising performance of the unpaired method. More details provided in Sec. 4.3 .

Figure 3: Validation of Expert Clinic Preference Alignment via Spearman’s correlation coefficient (

r

) between the medical experts (i.e., Experts Protocol Evaluation) and other tasks. Notably, our multi-dimensional evaluations (e.g., Full-Reference, No-Reference) demonstrated stronger correlation.

Figure 4: Illustration of the limitations of SDE-based methods. (A) and (B) show T-SNE [ 34 ] visualizations of features from synthetic high-quality images $\hat{\mathbf{y}}_{1}^{t_{i}}$ (blue points) and true high-quality images $\mathbf{y}_{2}$ (green points), extracted using the RETFound [ 40 ] and Ret-Clip [ 8 ] image encoders, respectively. Yellow arrows indicate the squeezing effect, while gray arrows denote feature drift over time. $d_{i}$ and $FID_{i}$ represent the center distance and corresponding FID score at time step $t_{i}$ . (C) visualizes attention drift in the skip-connection features maps of the generator [ 6 ] as the time step increases. Red boxes highlight lesion structures that gradually receive less attention. More details are discussed in Sec. 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 outlined below. (C) shows histograms for real-world No-Reference experiments after resampling. (D) presents reject-quality samples (e.g., overprocessed images).

Appendix figures & tables5 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 6: An illustrative medical expert clinical preference evaluation between (a) lesion preserving, (b) background preserving, and (c) structure-preserving.

Figure 7: Illustration of the Denoising Evaluation on the EyeQ dataset. The first and second columns show the high- and low-quality image references, respectively, while the remaining columns display the synthetic high-quality images generated by all baseline models.

Figure 8: Illustration of the Denoising Generalization Evaluation on the DRIVE and IDRID datasets. The first and second columns show the high- and low-quality image references, respectively, while the remaining columns display the synthetic high-quality images generated by all baseline models.

Figure 9: Illustration of Vessel and Lesion (EX and HE) Segmentation Experiments. The first column shows the reference segmentation masks, while the remaining columns display the segmentation results produced by all baseline models.

Figure 10: Illustration of the denoising results in the No-Reference Quality Assessment Experiments. The first column shows the input low-quality image, while the remaining columns display the synthetic high-quality images generated by all unpaired baseline models.

Explore similar work

FOCUS: Benchmarking Retinal Model Generalization from Foundation Vision Encoders to Multimodal LLMs
Sep 27, 2026David Restrepo, Chenwei Wu, Luis Filipe Nakayama +4Medical Image Analysis Medical Image Benchmarks
CentraleSupélec, Université Paris-Saclay, France · Cancer Data Science Unit, IHU PRISM, Université Paris-Saclay, CentraleSupélec, Gustave Roussy, INSERM, France · University of Michigan, USA +4
Luminosity-Adaptive Contrast Enhancement Using CLAHE for Retinal Fundus Images with Quantitative Validation and Comparative Analysis
Jul 20, 2026K. Mithra, Prem Kumar SanthanamImage Enhancement Fundus Imaging
Independent Researcher, Department of Information and Computer Engineering, Scottsdale, AZ, USA · Seidenberg School of Computer Science and Information Systems, Pace University, New York City, NY, USA
Ordinal Diffusion Models for Color Fundus Images
Feb 27, 2026Gustav Schmidt, Philipp Berens, Sarah MüllerDiffusion-Based Image Generation Diffusion Models
Hertie Institute for AI in Brain Health, University of Tübingen, Germany

Sep 27, 2026J/K move · Enter open · S save

FOCUS: Benchmarking Retinal Model Generalization from Foundation Vision Encoders to Multimodal LLMs

David Restrepo, Chenwei Wu, Luis Filipe Nakayama, Miguel L. Martins+3

CentraleSupélec, Université Paris-Saclay, France · Cancer Data Science Unit, IHU PRISM, Université Paris-Saclay, CentraleSupélec, Gustave Roussy, INSERM, France · University of Michigan, USA+4

Medical Image Analysis Medical Image Benchmarks Cross-Dataset Generalization Medical Image Classification VLM Evaluation

Open paper AskPDF arXiv

Progress in AI-based retinal image analysis has advanced with foundation models, yet evaluating their reliability remains challenging. Performance reported on a single dataset does not capture how models behave under dataset shift, across clinical definitions, or for different patient subgroups. This limitation is particularly critical in medical imaging analysis, where robustness, calibration, and fairness are essential for safe deployment. We introduce FOCUS (Foundation Ophthalmic Cross-Dataset Understanding under Shift), a cross-dataset benchmark for evaluating retinal fundus models that considers vision-only encoder models (VM), vision-language dual-encoder models (VLM), and multimodal large language models (MLLM). FOCUS harmonizes binary diabetic retinopathy, referable diabetic retinopathy, and glaucomatous optic neuropathy tasks across ten public datasets spanning diverse geographies, acquisition conditions, and label protocols. The benchmark evaluates models through a unified analysis layer that measures ranking performance, calibration, subgroup disparities, and image-quality robustness. We present a large-scale evaluation covering 532 base configurations and 228 MLLM configurations adapted through supervised fine-tuning with low-rank adaptation (LoRA). Results show that no model family consistently dominates across tasks and datasets: general VM encoders achieve the strongest average ranking performance, medical MLLMs are competitive but variable, and dual encoder VLMs benefit substantially from lightweight adaptation. Fine-tuning improves in-domain performance but exhibits heterogeneous transfer to external datasets, particularly in calibration. These findings demonstrate that retinal model evaluation is inherently multidimensional. FOCUS provides a practical framework and public benchmark to assess generalization, reliability, and robustness beyond single-dataset leaderboards

Abstract

Explore similar work

FOCUS: Benchmarking Retinal Model Generalization from Foundation Vision Encoders to Multimodal LLMs

Luminosity-Adaptive Contrast Enhancement Using CLAHE for Retinal Fundus Images with Quantitative Validation and Comparative Analysis

Ordinal Diffusion Models for Color Fundus Images