Fundus-Engine: Integrated Retinal Imaging Systems

• Fundus-Engine is a systems concept that couples retinal image acquisition, degradation modeling, enhancement, and diagnostic inference into an integrated pipeline.
• It employs modular designs such as enhancement-restoration, hardware adaptation, and quality assessment to improve clinical image fidelity.
• The framework incorporates multimodal diagnostic models and reasoning elements to provide comprehensive, reliable retinal evaluations.

Fundus-Engine is an umbrella designation for modular systems that acquire, assess, enhance, restore, and analyze retinal fundus images. In the cited literature, the term is applied to physically informed enhancement pipelines, portable imaging hardware with on-device reconstruction, quality-assessment and triage stacks, multimodal diagnostic models, and reasoning-oriented multimodal LLMs. This suggests that Fundus-Engine is best understood not as a single architecture but as a systems concept: a fundus-centered pipeline in which image formation, degradation modeling, restoration, diagnostic inference, and auditability are treated as coupled components rather than isolated tasks (Shen et al., 2020, Li et al., 2023, Rossi et al., 2022, Simmerer et al., 2024).

1. System concept and representative forms

A useful way to read the Fundus-Engine literature is to separate acquisition, image conditioning, and downstream analysis layers while recognizing that several papers explicitly couple them. Enhancement-oriented engines typically begin with quality assessment or degradation characterization, then apply restoration, and finally expose auxiliary outputs such as vessel masks, optic disc/cup masks, or artifact maps for downstream algorithms. Hardware-oriented engines instead shift part of the burden upstream by expanding field of view, suppressing glare optically, or replacing mechanical focusing with computational refocusing. Analysis-oriented engines use fundus-native representations, multimodal fusion, or reasoning traces to convert the corrected image into a disease label, severity grade, or textual explanation (Shen et al., 2020, Rossi et al., 2022, Zun et al., 25 Jun 2025, Deng et al., 9 Apr 2026).

Category	Representative systems	Reported role
Enhancement/restoration	cofe-Net (Shen et al., 2020), GFE-Net (Li et al., 2023), SCR-Net (Li et al., 2022), FD3 (Kim et al., 2024), PTL multi-pass restoration (Phan et al., 14 Apr 2025)	Correct low-quality, cataract-degraded, or mixed-degradation fundus images
Acquisition hardware	Portable HDR widefield camera (Rossi et al., 2022), diffuser-based computational camera (Simmerer et al., 2024)	Improve contrast, field of view, glare suppression, or refocusing at capture time
Quality assessment	FundaQ-8 (Zun et al., 25 Jun 2025)	Assign interpretable quality scores and route retake/proceed decisions
Diagnostic inference	SwinECAT (Gu et al., 29 Jul 2025), Image+Fundus (Jang et al., 2024), CFT (Xiao et al., 2024), UOPSL (Zhao et al., 10 Sep 2025)	Perform disease classification, grading, or fundus-only inference with multimodal priors
Reasoning and reporting	Fundus-R1 (Deng et al., 9 Apr 2026)	Generate knowledge-aware answers and reasoning traces from public data

2. Imaging physics, degradation, and acquisition-side engines

A central theme in Fundus-Engine design is that retinal image degradation is not treated as generic photographic corruption. The ophthalmoscope-aware model used with cofe-Net explicitly attributes low quality to shared pupil illumination and imaging, inter-reflections, stray light leakage, aperture and dynamic-range limits, defocus, eyeball motion, and undesired objects such as dust or grains. Its degradation model separates non-uniform illumination, PSF blur, and additive artifacts. Representative formulations are

$$I_{\text{light}} = \operatorname{clip}\!\big(\alpha (J \cdot G_L + x) + \beta; s\big),$$

$I_{\text{blur}} = x * h_B + n,$

and

$$I_{\text{art}} = x + \sum_{k=1}^{K} G_R(r_k/4,\sigma_k)\cdot o_k,$$

with synthetic parameter ranges chosen to emulate underexposure, leakage, blur, noise, and artifact density observed in clinical fundus imaging (Shen et al., 2020).

Acquisition-side Fundus-Engines pursue the same objective by changing the camera. A portable nonmydriatic widefield fundus camera combines miniaturized indirect ophthalmoscopy illumination, orthogonal polarization control, and flash-bracketed HDR fusion. It achieves a snapshot field of view of $101^\circ$ eye-angle ($67^\circ$ visual-angle), expandable to $190^\circ$ eye-angle ($134^\circ$ visual-angle) horizontally and $146^\circ$ eye-angle ($100^\circ$ visual-angle) vertically using a fixation target, while HDR fusion preserves detail in both bright and dark retinal regions. The system was evaluated under ISO 10940:2009 constraints, with retinal radiant exposure per acquisition reported as $4.71 \times 10^{-6}\,\mathrm{J/cm^2}$, well below the allowed exposure (Rossi et al., 2022).

A more radical capture architecture replaces mechanical focusing with a pupil-conjugate holographic diffuser. In that design, the sensor records a caustic PSF whose shape varies predictably with refractive error, and post-capture Tikhonov-regularized deconvolution performs computational refocusing across at least $I_{\text{blur}} = x * h_B + n,$0 diopters. The demonstrated field of view is at least $I_{\text{blur}} = x * h_B + n,$1, approximately $I_{\text{blur}} = x * h_B + n,$2 in vivo, with measured resolution of $I_{\text{blur}} = x * h_B + n,$3–$I_{\text{blur}} = x * h_B + n,$4 line pairs per mm across defocus. This is lower than a properly focused conventional camera but removes pre-capture focus adjustment and thereby shifts complexity from mechanics to calibration and reconstruction (Simmerer et al., 2024).

3. Enhancement and restoration architectures

Enhancement-oriented Fundus-Engines differ mainly in where they place prior knowledge. cofe-Net introduces a three-branch architecture with Low-Quality Activation, Retinal Structure Activation, and a two-scale encoder-decoder correction module. LQA produces an artifact attention map $I_{\text{blur}} = x * h_B + n,$5 that modulates low-level correction features through $I_{\text{blur}} = x * h_B + n,$6, while RSA uses a fundus-pretrained ResNet-34 to preserve vessels and optic disc/cup structure. Its total loss combines reconstruction, artifact-mask supervision, artifact-focused reconstruction, and RSA segmentation supervision with default weights $I_{\text{blur}} = x * h_B + n,$7, $I_{\text{blur}} = x * h_B + n,$8, and $I_{\text{blur}} = x * h_B + n,$9. On synthesized degraded DRIVE/Kaggle images, the full model reports PSNR $$I_{\text{art}} = x + \sum_{k=1}^{K} G_R(r_k/4,\sigma_k)\cdot o_k,$$0 and SSIM $$I_{\text{art}} = x + \sum_{k=1}^{K} G_R(r_k/4,\sigma_k)\cdot o_k,$$1, and the corrected images improve CE-Net vessel segmentation from AUC $$I_{\text{art}} = x + \sum_{k=1}^{K} G_R(r_k/4,\sigma_k)\cdot o_k,$$2 to $$I_{\text{art}} = x + \sum_{k=1}^{K} G_R(r_k/4,\sigma_k)\cdot o_k,$$3 and M-Net optic disc/cup segmentation from F-score $$I_{\text{art}} = x + \sum_{k=1}^{K} G_R(r_k/4,\sigma_k)\cdot o_k,$$4 to $$I_{\text{art}} = x + \sum_{k=1}^{K} G_R(r_k/4,\sigma_k)\cdot o_k,$$5 (Shen et al., 2020).

GFE-Net replaces explicit structure labels with frequency self-supervised representation learning. It trains a shared encoder and two decoders: $$I_{\text{art}} = x + \sum_{k=1}^{K} G_R(r_k/4,\sigma_k)\cdot o_k,$$6 reconstructs high-frequency maps from degraded views, and $$I_{\text{art}} = x + \sum_{k=1}^{K} G_R(r_k/4,\sigma_k)\cdot o_k,$$7 restores RGB images while receiving decoded features from $$I_{\text{art}} = x + \sum_{k=1}^{K} G_R(r_k/4,\sigma_k)\cdot o_k,$$8 at every level. The key operator is a Gaussian high-pass filter,

$$I_{\text{art}} = x + \sum_{k=1}^{K} G_R(r_k/4,\sigma_k)\cdot o_k,$$9

used both as supervision and as the basis for representation learning. Because vessels and lesions are concentrated in high frequencies, this design is intended to learn degradation-robust, structure-aware features without paired labels or test-domain access. Trained only from 40 clear DRIVE images with synthesized degradations, GFE-Net reaches SSIM $101^\circ$0 and PSNR $101^\circ$1 on FIQ, SSIM $101^\circ$2 and PSNR $101^\circ$3 on RCF, and runs at $101^\circ$4 GMac with $101^\circ$5 s per $101^\circ$6 image (Li et al., 2023).

Cataract-specific restoration emphasizes structure consistency. SCR-Net constructs a synthesized cataract set in which each degraded image shares identical anatomy with its clean source, then constrains restoration through high-frequency component alignment. HFCs are extracted by subtracting a Gaussian low-pass image, and the network jointly optimizes HFC alignment, image reconstruction, and HFC cycle-consistency. On the RCF dataset it reports SSIM $101^\circ$7, PSNR $101^\circ$8, vessel-segmentation IoU $101^\circ$9, FIQA $67^\circ$0, diagnostic F1 $67^\circ$1, and Cohen’s kappa $67^\circ$2, outperforming CycleGAN, CofeNet, I-SECRET, and other baselines in the reported comparison (Li et al., 2022).

FD3 adopts a direct diffusion bridge rather than GAN-style or purely feed-forward restoration. Its bridge is

$67^\circ$3

with $67^\circ$4 defined as a CLAHE-preferred target and $67^\circ$5 synthesized through a clinically guided forward model. The network learns $67^\circ$6 by minimizing an $67^\circ$7 objective over random $67^\circ$8, and inference deterministically updates

$67^\circ$9

Using 10 function evaluations, FD3 reports on EyeQ with the proposed forward model PSNR $190^\circ$0, FID $190^\circ$1, and vessel-segmentation IoU $190^\circ$2; on FPE it reports PSNR $190^\circ$3, FID $190^\circ$4, and IoU $190^\circ$5. In a 50-image in-vivo clinician ranking study, it achieves $190^\circ$6, ahead of PCE-Net and CLAHE (Kim et al., 2024).

Progressive multi-pass restoration places the prior in iteration. The PTL framework trains a CycleGAN for blind restoration and then reuses the best pass as initialization for subsequent passes, progressively refining the restored outputs. On DeepDRiD, downstream DR screening after three passes reaches Accuracy $190^\circ$7, Precision $190^\circ$8, Sensitivity $190^\circ$9, and F1 $134^\circ$0, compared with $134^\circ$1, $134^\circ$2, $134^\circ$3, and $134^\circ$4 on the original images (Phan et al., 14 Apr 2025).

4. Quality assessment, gating, and safety

A mature Fundus-Engine does not enhance every image indiscriminately. FundaQ-8 formalizes image quality as eight clinically interpretable parameters—Resolution, Field of View, Color Fidelity, Presence of Artifacts, Vessels, Macula, Optic Disc, and Optic Cup—each scored on a $134^\circ$5–$134^\circ$6 Likert scale. With equal weights, the normalized composite score is

$134^\circ$7

and the operational thresholds are $134^\circ$8 for Bad, $134^\circ$9 for Medium, and $146^\circ$0 for Good. Recommended hard-fail rules include Coverage $146^\circ$1, OD $146^\circ$2, Macula $146^\circ$3, Artifacts $146^\circ$4, or simultaneous Vessels $146^\circ$5 and Resolution $146^\circ$6. On the internal test set, the ResNet18 regressor reports MSE $146^\circ$7, MAE $146^\circ$8, RMSE $146^\circ$9, and $100^\circ$0; on EyeQ, predicted $100^\circ$1 shows Spearman $100^\circ$2 against the Good/Usable/Rejected label scale (Zun et al., 25 Jun 2025).

The quality gate is also a safety mechanism. Multiple enhancement papers explicitly warn that restoration should remain a preprocessing step rather than a substitute for clinical interpretation. Reported risks include over-suppression of subtle lesions, smoothing of microaneurysms, hallucinated structure under aggressive enhancement, and failure under extreme blur or occlusion. The proposed safeguards are consistent across the literature: preserve vessel and disc/cup edges through explicit structure priors, monitor lesion candidate counts before and after enhancement, expose confidence measures or soft masks, and review original and corrected images side-by-side when subtle pathology is suspected (Shen et al., 2020, Li et al., 2022).

A second safety issue is domain shift. Synthetic degradations can cover illumination, haze, blur, noise, and artifact spectra, but several papers state that new cameras or optics may still require fine-tuning, site-specific parameter calibration, or adaptation modules. This is particularly explicit in generic enhancement and camera-adaptation work, and it implies that Fundus-Engine deployment should include device logging, periodic quality audits, and a mechanism to detect when restoration or diagnosis has drifted outside its validated operating regime (Li et al., 2023, Lin et al., 2022).

5. Downstream diagnosis, multimodal fusion, and domain adaptation

At the classification layer, SwinECAT exemplifies a transformer-based diagnostic Fundus-Engine. It combines a four-stage Swin Transformer backbone with Efficient Channel Attention after each stage, using stage depths $100^\circ$3. On the EDID dataset of 16,140 fundus images and 9 classes, it reports Accuracy $100^\circ$4, Macro F1 $100^\circ$5, Weighted F1 $100^\circ$6, and $100^\circ$7M parameters, improving over the baseline Swin Transformer by $100^\circ$8 percentage points in accuracy and $100^\circ$9 in Macro F1 (Gu et al., 29 Jul 2025).

Multimodal engines use cross-modal correspondence rather than simple voting. The Cross-Fundus Transformer processes paired CFP and IFP images through dual ViT streams and fuses patch tokens with symmetric Cross-Fundus Attention. It keeps modality-specific class tokens for single-modality supervision and fuses cross-attended patch representations by elementwise max pooling. On 1,713 paired CFP/IFP images for five-grade DR classification, it reaches QWK $4.71 \times 10^{-6}\,\mathrm{J/cm^2}$0, Accuracy $4.71 \times 10^{-6}\,\mathrm{J/cm^2}$1, and Macro-F1 $4.71 \times 10^{-6}\,\mathrm{J/cm^2}$2, outperforming single-modality ViTs and multi-modal baselines such as voting-average and feature-concat (Xiao et al., 2024).

A different form of multimodality appears in UOPSL, which uses unpaired OCT data to learn an OCT-space predilection sites matrix $4.71 \times 10^{-6}\,\mathrm{J/cm^2}$3 and then removes the OCT branch at inference time. Disease text descriptions bridge fundus and OCT during pretraining, and at fundus-only inference the learned $4.71 \times 10^{-6}\,\mathrm{J/cm^2}$4 is used as a query in cross-attention against class-specific text. This allows a fundus-only classifier to retain OCT-derived spatial priors. Across nine evaluation datasets, UOPSL reports, for example, AUROC/PRC $4.71 \times 10^{-6}\,\mathrm{J/cm^2}$5 on IDRID, $4.71 \times 10^{-6}\,\mathrm{J/cm^2}$6 on GF, and $4.71 \times 10^{-6}\,\mathrm{J/cm^2}$7 on JSIEC, with strong zero-shot performance on unseen categories (Zhao et al., 10 Sep 2025).

Foundation-style pretraining offers another route to reusable fundus representations. The Image+Fundus model pretrains a ResNet-50 first on ImageNet and then on 113,645 institutional fundus images for abnormality detection. In internal abnormality classification at $4.71 \times 10^{-6}\,\mathrm{J/cm^2}$8 px, linear probing reaches AUC $4.71 \times 10^{-6}\,\mathrm{J/cm^2}$9, and external validation reaches AUC $I_{\text{blur}} = x * h_B + n,$00 on JSIEC and $I_{\text{blur}} = x * h_B + n,$01 on RFMiD. In multi-disease internal evaluation, the same backbone reports F1 $I_{\text{blur}} = x * h_B + n,$02 for AMD, $I_{\text{blur}} = x * h_B + n,$03 for pathologic myopia, and $I_{\text{blur}} = x * h_B + n,$04 for ERM, though the transfer to vessel segmentation is reported as weaker than plain ImageNet pretraining (Jang et al., 2024).

Domain adaptation becomes critical when the acquisition device changes. For CVD risk estimation, a camera-adaptive engine combines cross-laterality feature alignment on UK Biobank with a self-attention camera adaptor trained on paired Topcon/Mediwork fundus images. The best configuration, Weight2 + CLFA + SACA, reports $I_{\text{blur}} = x * h_B + n,$05 and post-adaptation cross-camera $I_{\text{blur}} = x * h_B + n,$06, outperforming DAN, DANN, MDD, Pix2Pix GAN, and CycleGAN in the reported comparisons (Lin et al., 2022).

A common misconception is that larger fundus foundation models necessarily dominate fine-grained tasks. In DME detection, this is not supported uniformly: EfficientNet-B0 ranked first or second in most settings, RETFound showed its strongest behavior on OEFI, and FLAIR’s most competitive behavior appeared in carefully prompted zero-shot mode. This indicates that foundation-scale pretraining is task-dependent and does not eliminate the continued relevance of lightweight CNN baselines for localized lesion detection (Arellano et al., 8 Oct 2025).

6. Reasoning-scale engines, deployment, and reproducibility

The most expansive Fundus-Engine formulation is reasoning-centric. Fundus-R1 treats CFP, OCT, and UWF interpretation as a vision-language problem and trains a fundus-reading MLLM exclusively from public data. The system assembles label- and modality-conditioned knowledge from EyeWiki, AAO guidance, and PMC, extracts image-specific findings with a generic MLLM, composes knowledge-aware reasoning traces, filters them with an LLM judge, and then trains with supervised fine-tuning followed by RLVR with a process reward. The public-only corpus contains 168,938 images, more than 94% with image-level labels, and yields 80,115 retained reasoning traces from 146,425 synthesized traces. Relative to generic Qwen2.5-VL-3B, Fundus-R1-3B improves the average benchmark score from $I_{\text{blur}} = x * h_B + n,$07 to $I_{\text{blur}} = x * h_B + n,$08, and Fundus-R1-7B reaches $I_{\text{blur}} = x * h_B + n,$09 across FunBench, Omni-Fundus, and GMAI-Fundus (Deng et al., 9 Apr 2026).

Deployment-oriented work extends the engine concept across domains rather than modalities alone. For example, semi-supervised adaptation from traditional fundus to ultra-wide-field images uses CycleGAN-based image translation, pseudo-labeling, consistency regularization, and MixUp. On UWF test data, the best reported setting reaches Accuracy $I_{\text{blur}} = x * h_B + n,$10, Precision $I_{\text{blur}} = x * h_B + n,$11, Recall $I_{\text{blur}} = x * h_B + n,$12, Specificity $I_{\text{blur}} = x * h_B + n,$13, and F1 $I_{\text{blur}} = x * h_B + n,$14, outperforming UWF-only training, fine-tuning, knowledge distillation, pseudo-labeling alone, and MixUp alone (Ju et al., 2020).

Across these strands, reproducibility is uneven but improving. Public repositories are explicitly reported for cofe-Net, SCR-Net, FD3, and the CVD camera-adaptation system, and code release is stated for several other systems. Training stacks are most often PyTorch-based, with reported configurations including V100 or A5000-class GPUs for enhancement models, Quadro RTX 6000 for FundusGAN, and 8× H800 GPUs for Fundus-R1. A consistent deployment pattern also emerges: ingest image and metadata, assess quality, optionally characterize degradation, restore or enhance only if warranted, expose auxiliary masks or attention maps, route to a downstream task model, and log intermediate outputs for audit. This suggests that the enduring significance of Fundus-Engine lies less in any single architecture than in the insistence that capture physics, quality control, restoration, multimodal evidence, and diagnostic reasoning be assembled into a traceable clinical pipeline rather than optimized as disconnected benchmarks (Shen et al., 2020, Kim et al., 2024, Deng et al., 9 Apr 2026).