Early automated systems extracted hand-crafted descriptors vessel maps, microaneurysms, haemorrhages, and exudates from colour fundus images and applied SVMs, kNN, or Bayesian models to estimate DR severity. These pipelines required heavy preprocessing, were sensitive to illumination and device shifts, and often struggled on heterogeneous cohorts [16]. To alleviate data scarcity, works explored contrast enhancement and GAN-based augmentation to stabilise prognostication [17], but such strategies still operated on fundus-only inputs and lacked direct angiographic signals unless dye-based FA or OCTA was acquired [18]. Robust learning under imperfect labels (loss correction and noise-aware training) was also studied [19], and objective image-quality metrics such as SSIM, PSNR, MSE, and FSIM became standard for screening inputs and evaluating restorations [20].

Modern screening is dominated by convolutional and transformer-based architectures, spanning segmentation (vessel/lesion localisation) and end-to-end grading [21]. Lesion aware and multi-loss designs improved microvasculature delineation and robustness [22], while hybrid pipelines with vision transformers strengthened representation learning for synthesis and predict workflows [23]. Transfer learning and representation pretraining trends (including vision–language supervision) further enhanced feature reuse across cohorts [24]. Persistent challenges include large annotation demands, class imbalance, and domain shift [25]. Diffusion models recently emerged as strong priors for denoising and augmentation on retinal imagery, improving fidelity and training stability [26]. Nevertheless, most pipelines are fundus-only and cannot directly model nonperfusion or leakage patterns that are best visualised on FA or OCTA; studies using UWF-FA demonstrate grading value but still require invasive dye [27].

Cross-modal retinal synthesis aims to recover angiography-like information from non-invasive images. Foundational conditional adversarial frameworks established paired image-to-image translation, and ophthalmic variants adapted architectures, objectives, and priors to encode retinal structure [28]. UWAT-GAN introduced an ultra-wide-angle, multi-scale generator to map UWF-RI to UWF-FA, but relied on only a few dozen matched pairs, limiting generalisation and phase modelling [29]. UWAFA-GAN increased data scale and integrated attention and registration enhancement to sharpen vasculature fidelity; however, it produced a single venous-phase frame and did not quantify downstream DR stratification gains. Other directions focused on targeted synthesis or controllability: lesion-centric DR-LL-GAN, Wasserstein-based retinal synthesis, controllable StyleGAN variants, and high-fidelity semantic manipulation. Some explored grading via the generative prior itself, but comprehensive multi-phase UWF-FA generation coupled with an integrated classifier and validated on downstream clinical tasks remains uncommon [30].

Beyond ophthalmology, adversarial and diffusion paradigms matured for translation and restoration. Unsupervised adversarial diffusion enabled cross modality mapping without dense pairing [31], while image-to-image diffusion provided stable, controllable synthesis [32]. Disentangled and regularised GANs promoted factorised latent structure and interpretable controls [33]; surveys chart diffusion advantages in likelihood training, diversity, and perceptual quality [34]. At the same time, the community cautioned against overreliance on single realism metrics, advocating multi metric assessment [35] and task based validation principles directly relevant to retinal synthesis and evaluation [36]. Recent hybrid generative–classification frameworks have further demonstrated the synergy between image synthesis and diagnostic modeling in diabetic retinopathy [37]. Gencer et al. (2025) combined GAN-based augmentation with denoising autoencoders and an EfficientNet-B0 classifier, achieving 99% accuracy, recall, and specificity on a high-resolution OCT dataset [38].

UWF imaging expands coverage to the periphery, where clinically significant lesions and nonperfusion can alter staging and management. Widefield OCTA studies link nonperfusion areas with DR severity and progression, underscoring the value of angiographic signals beyond colour fundus. Deep-learning pipelines trained directly on UWF-FA demonstrate strong grading potential, yet FA requires dye injection, with workflow implications and rare adverse reactions. These factors motivate non-invasive synthesis that preserves diagnostic utility while improving accessibility [39].

In summary, existing retinal synthesis studies often train on limited paired data, restrict outputs to a single angiographic phase (Table 1), or focus on narrow fields and lesion slices; few quantify downstream diagnostic gains with an integrated classifier. Registration errors and domain shift further hinder generalisation at UWF resolution [40]. Our work addresses these gaps with a large paired UWF-RI and UWF-FA corpus, a phase aware generator that synthesises initial, mid, and final FA frames from a single UWF-RI input, and an integrated Swin Transformer classifier; we evaluate both multi metric image quality and downstream DR stratification, linking image realism to clinical performance while avoiding dye related risks.

We used two datasets that play complementary roles in development and external validation. The primary ultra-widefield (UWF) cohort comprising 1198 patients (2915 UWFRI, 17,854 UWFFA) is an internally curated research dataset collected under institutional ethics approval, used for all phase-aware synthesis and evaluation experiments. Public datasets such as ODIR and Messidor-2 are referenced only for auxiliary pretraining and external classification validation; neither contains UWF-FA images. This clarification resolves earlier inconsistencies in dataset identity and ensures reproducibility. All UWF images used for training and testing were acquired using the same ultra-widefield imaging platform and optical configuration to ensure consistency in field curvature, illumination, and device characteristics, cross-vendor generalization was not within the current study’s scope.

Primary development dataset (ODIR). For training and tuning the cross-modal generator, we used 3800 ultra widefield retinal images UWF_RI and 1100 ultra widefield fluorescein angiography images UWF_FA. Images were captured at 3500 × 3500 pixels with an approximately 180∘ field of view and de-identified prior to analysis [48].

Angiographic timing was organized in three phases for UWF_FA: initial (20–50 s post-injection), mid (1–4 min), and final (>4 min). Unless noted, training emphasized post-venous frames, while validation and testing covered all phases.

Frames with severe eyelid/iris occlusion, motion blur, or illumination artifacts were excluded during quality review. For paired UWF_RI–UWF_FA sequences (same eye and visit), only pairs passing stringent registration (see Section 4) were retained. Pairs with Dice below 0.6 were discarded, removing approximately 7.3% of candidates. All data were de-identified and used solely for research purposes.

Patient-level splitting (3:1:1 train:validation:test) was utilized to avoid leakage, maintaining DR class balance and FA phase balance. The same policy was used for Messidor-2. All test results reflect a single held-out set with no overlap with training.

Preprocessing included mask-based cropping to suppress periocular artifacts, resolution harmonization, intensity normalization to [0, 1], and optional histogram matching for cross-session consistency. For generation, UWF_RI and UWF_FA were resized to 1080 × 1080; for classification, images were resized to 1024 × 1024. Transformations were applied identically within a paired sample to preserve geometry across modalities and phases.

Training was performed on three phase specific generators (initial, mid, final) to produce UWF_FA from UWF_RI. Data augmentation included random resized cropping (scale 0.4–3.0), horizontal/vertical flips, and random rotations within 0–30 degrees. Each generator trained for 45 epochs (batch size 6, learning rate 10−4). Settings were selected empirically for stable convergence given available compute. To mitigate overfitting risk associated with limited batch size and high model capacity, several regularization strategies were employed. Instance normalization and dropout layers (p = 0.3) were integrated within the generator to discourage co-adaptation. Early stopping was applied based on validation MAE, and model checkpoints were selected at minimum validation loss. Additionally, extensive data augmentation—spatial flips, rotations, scaling, and intensity jitter—was used to increase diversity and reduce memorization bias across the three phase-specific generators.

To quantify downstream value of synthetic angiography, a Swin Transformer with an MLP head was trained on four compositions:

Set A: Real UWF_RI only (baseline).

All runs used identical splits, ImageNet initialization, and class balancing. Swin features were reduced to 1024-dimensional embeddings and passed to a fully connected layer with softmax. We used Adam (learning rate (1 ×10−5), batch size 4) with early stopping on validation AUC.

The Gradient Variance Loss (GVL) term encourages local edge consistency and is defined as: ℒGVL=1N∑i=1N(Var(∇xIi)+Var(∇yIi)),where Var denotes the variance of spatial gradients.

For synthesis, we report MAE, PSNR, SSIM, and MS-SSIM between generated and ground-truth UWF_FA on the held-out test set. For DR classification, we report AUC, APR, F1, sensitivity, specificity, and accuracy, indicating statistical significance relative to Set A.

This subjective assessment followed a double-blind, within-subject design where experts were unaware of image origin (synthetic vs. real). Two ophthalmologists independently rated 50 test pairs across FA phases on a 5-point quality scale (1 best, 5 worst). Inter-rater reliability used Cohen’s weighted kappa. A Turing-style task asked readers to label 25 FA images as real or synthetic; a third adjudicator resolved disagreements. The resulting kappa of 0.74 indicates substantial agreement. In total, 50 unique UWF-FA test pairs were presented per reader across three angiographic phases (initial, mid, final), including both real and synthesized samples balanced in equal proportion. For the Turing-style evaluation, each ophthalmologist reviewed 25 randomly sampled FA images—12 synthetic and 13 real—under blinded conditions. Images were classified as “synthetic” if they exhibited any atypical vessel branching, inconsistent dye diffusion, or unnatural texture patterns not characteristic of genuine angiograms. These predefined visual cues, established prior to reading, ensured consistent interpretation criteria across experts.

Accurate alignment between UWFRI and UWFFA images is critical for supervised translation and for preserving lesion geometry across modalities. Vessel maps S(⋅) are extracted using the Retina-based Microvascular Health Assessment System (RMHAS) to emphasize stable vasculature and suppress illumination artifacts [50]. Keypoints and descriptors are computed using AKAZE feature detection [51] and tentative correspondences are refined through RANSAC-based homography estimation [52] to yield the warp matrix H∈R3×3. Pairs failing geometric sanity checks (absolute rotation angle >2.5 radians or scale ratio outside [0.7,1.5]) are discarded, and only those achieving a Dice overlap coefficient ≥0.6 between warped vessel maps are retained for training. This registration process yields geometrically aligned image pairs {(x,y)}, where x denotes the UWFRI image and y the corresponding phase-specific UWFFA.

Separate generators (G∧init, G∧mid, G∧final) are trained to map UWF_RI to each FA phase. Each generator follows a high-resolution conditional GAN backbone inspired by pix2pixHD with multi scale discriminator(s) to stabilize training on UWF fields. For a given phase (p∈{init,mid,final}), the generator produces y^p=Gp(x) and the discriminator Dp distinguishes the real pair (x,yp) from the synthetic pair (x,y^p).

To quantify diagnostic utility of synthetic angiography, integrates a Swin Transformer classifier that fuses real UWF_RI with zero to three synthetic UWF_FA phases (Sets A–D in Section 3). Each available image (RI or FA phase) is fed through a shared Swin backbone f(⋅) to obtain a pooled embedding e=f(⋅)∈Rd. Embeddings from the available modalities are concatenated and projected to a 1024-dimensional representation, which is classified with a fully connected layer and softmax. Formally, for a sample with modalities ℳ⊆{RI,FAinit,FAmid,FAfinal}, (7)E = ‖m∈ℳf(x(m)) ∈ R|ℳ|d,z = MLP1024(E),c^ = softmax(Wz+b),and the cross-entropy; (8)ℒcls = −∑k=1K1[c=k] log⁡c^kis minimized over DR classes k∈{0,…,4}. This late-fusion design allows ablations over modality availability (Sets A–D) without modifying backbone internals.

At test time, a single UWF_RI image is mapped by the three phase-specific generators to produce (y^_init,y^_mid,y^_final). The classifier consumes the real UWF_RI and the selected synthetic phases according to the target setting (A–D) to output DR stage probabilities and the final prediction. This design enables deployment with or without synthetic FA, and permits phase-wise sensitivity analysis of diagnostic contributions.

Generators and discriminators follow common residual and multi-scale design choices for high-resolution translation. Training uses the protocols in Section 3 (phase-specific models; 45 epochs; batch size 6; learning rate 0.00015; augmentations: random resized crop, flips, rotations). The classifier employs ImageNet initialization, Adam optimization with learning rate (10−5), batch size 4, and early stopping on validation AUC. All transforms within a paired sample are applied identically across modalities to preserve alignment.

After quality screening and registration filtering, the final cohort comprised 1198 patients contributing 2915 UWF_RI images and 17,854 UWF_FA images. On average, each subject contributed approximately two UWF_RI frames and fifteen UWF_FA frames collected at a single clinical visit, ensuring phase consistency for paired supervision. Demographically, 689 patients (57.5%) were male, with a median age of 56.34 years (range: 44.12–66.89 years). By angiographic stage, the dataset contained 1927 initial, 9892 mid, and 6035 final phase pairings, reflecting distinct circulation states exploited by our phase-aware generators. The overall workflow for data preparation, model training, and evaluation is summarized in Fig. 3.

To enhance reproducibility across splits, all images underwent the same preprocessing sequence: normalization, optional histogram matching, resizing 512 × 512 for analysis figures; 1080 × 1080 for generator training; 1024 × 1024 for classification), and intensity scaling to [0, 1]. This uniform pipeline reduced device/session variability and preserved cross-modal geometry within pairs.

Per-phase reconstruction quality is reported, using MAE, PSNR, SSIM, and MS-SSIM on the held-out test pairs (Table 2). All fidelity metrics (MAE, PSNR, SSIM, MS-SSIM) were computed on the 8-bit [0–255] intensity scale prior to normalization, explaining the magnitude of MAE values near 100 despite normalized preprocessing. Differences reported in Table 2 are statistically significant at p<0.05 (two-tailed paired t-test). Trend plots for MAE, PSNR, and SSIM/MS-SSIM are shown in Figs. 4–6, respectively.

The mid stage attained the lowest MAE (Table 2), indicating the most faithful pixel-level reconstruction. A slight PSNR decrease from initial to final phase (Fig. 5) suggests modest noise or contrast dispersion later in circulation, yet absolute PSNR values remain high. SSIM values are stable (Fig. 6), while MS-SSIM dips in mid phase and rebounds in final, indicating transient loss of fine multi-scale structures during mid-phase contrast dynamics.

Two ophthalmologists achieved substantial inter-rater agreement (Cohen’s weighted κ = 0.78–0.82 across phases), confirming consistency of expert grading, as summarized in Table 3 and Fig. 7. In a Turing-style task, experts mislabeled 50%–70% of synthetic UWF_FA images as real, supporting strong perceptual realism. Mid-phase images yielded the highest agreement (kappa 0.82) despite slightly higher subjective variability (R2 mean 2.01), indicating that clinically salient structures remained consistently interpretable.

Diagnostic utility was quantified by training a Swin-based classifier on four dataset compositions (Sets A–D; see Section 3). Performance metrics (AUC, APR, F1, sensitivity, specificity, accuracy) are reported in Table 4 with visual summaries in Figs. 8 and 9. Adding synthetic FA improved all task metrics relative to baseline (Set A), with full multi-phase integration (Set D) achieving the best overall performance (AUC 0.910, APR 0.792, accuracy 0.829). Significance tests indicate these gains are unlikely due to chance (Table 4). The monotonic rise in sensitivity from Sets A to D suggests that angiographic cues—nonperfusion, leakage, peripheral vasculature—provide complementary information to UWF_RI.

Fig. 10 illustrates side-by-side comparisons of real and synthesized UWF_FA across initial, mid, and final phases for representative conditions (retinal macroaneurysm, normal retina, DR, and retinal vein occlusion). The synthesized images preserve macrovessel topology and lesion boundaries while reproducing phase-dependent contrast filling, supporting both perceptual realism and clinical interpretability.

Across quantitative metrics, expert review, and downstream classification, the proposed phase-aware synthesis yields high-fidelity UWF_FA that enhances DR stratification when combined with UWF_RI. Mid-phase reconstructions exhibit the lowest MAE; multi-scale structure is transiently reduced during mid-phase but recovers in final-phase views; and multi-phase synthetic integration delivers the strongest end-task performance.

Our study advances beyond earlier UWF translation approaches by scaling paired data, enforcing strict registration quality, and learning phase-specific generators. Prior systems such as UWAT-GAN focused on single venous-phase synthesis with limited paired samples, which constrains generalization and fails to model temporal angiographic dynamics. In contrast, we produce initial, mid, and final phases under a unified training and evaluation protocol, enabling richer depiction of nonperfusion and leakage patterns. This broader temporal coverage and the integration with a modern classifier strengthen the link between perceptual realism and clinical task performance.

UWF imaging captures peripheral pathology that can shift disease staging and management. Angiographic information remains the gold standard for visualizing capillary dropout and leakage but requires dye injection with workflow burden and rare adverse reactions. Our results show that synthetic UWF_FA can complement UWF_RI to improve DR stratification (Set D vs. Set A in Table 4), suggesting a practical path to safer, scalable screening and triage where invasive FA is unavailable or contraindicated.

The mid phase achieved the lowest MAE (Table 2), indicating the most faithful pixel-level reconstruction during peak circulation contrast. The slight PSNR decline from initial to final phase (Fig. 5) likely reflects dispersion and noise accumulation later in the sequence, yet absolute values remain high. SSIM is stable across phases (Fig. 6); the MS-SSIM dip in the mid phase suggests transient loss of very fine multi-scale structure during dynamic filling, which is restored by the final phase. These patterns are consistent with the known temporal physiology of FA and the trade-offs between global fidelity and micro-structure preservation at high resolution.

Two ophthalmologists provided consistent quality ratings with substantial agreement (κ 0.78–0.82; Table 3, Fig. 7), and mislabeled 50%–70% of synthetic images as real in a Turing-style task. These findings corroborate the perceptual validity of the synthesized angiograms and align with the quantitative fidelity metrics. Nevertheless, we emphasize that visual plausibility alone is insufficient; multi-metric assessment and task-based endpoints remain essential to guard against overfitting to any single realism measure. While the synthesized UWF-FA images exhibit high quantitative fidelity and strong clinical realism, they are intended to complement rather than replace conventional FA in diagnostic workflows. Synthetic FA generation offers a non-invasive, rapid, and low-risk alternative for scenarios where dye-based angiography is contraindicated or unavailable. In clinical decision-making, such images can support early screening, triage, and treatment planning, but final therapeutic decisions should remain based on physician review and, where appropriate, confirmatory FA acquisitions.

Integrating synthetic FA incrementally improved all classification metrics relative to a UWF_RI only baseline, with the full multi-phase setting (Set D) yielding the best AUC, APR, sensitivity, and accuracy (Table 4 and Figs. 8 and 9). The monotonic gain in sensitivity from Sets A to D indicates that angiographic cues provide complementary information to color imaging particularly for detecting nonperfusion and peripheral vascular pathology not fully captured by UWF_RI. This supports the design choice of phase-specific generators and late-fusion classification.

First, despite stringent registration (Section 4), peripheral distortion and residual artifacts can degrade synthesis near the edges of UWF fields, potentially affecting lesion depiction. Second, although we curated large paired data, domain shift across devices and sites may persist; prospective multi-center validation is needed. Third, Messidor-2 served as external data for classification with synthetic FA derived from UWF_RI, which may not fully reflect clinical deployment where heterogeneous imaging protocols are common. Fourth, our evaluation focused on DR; generalization to other retinal diseases (e.g., RVO, RAM) is suggested by qualitative examples (Fig. 10) but requires dedicated studies. Finally, reader studies included a limited number of experts; expanding to larger, geographically diverse panels will improve confidence intervals and reduce annotation bias.

Several directions can further raise clinical readiness. (i) Data and domain expansion: incorporate additional vendors, populations, and pathologies, with patient-level temporal follow-up to study progression. (ii) Robust learning: add domain adaptation and self-supervised pretraining to mitigate shift, and uncertainty estimation to flag unreliable synthesis. (iii) Architectural advances: explore diffusion or hybrid adversarial–diffusion objectives for sharper microvasculature while retaining stability. (iv) Lesion-aware evaluation: augment global metrics with vessel-wise and lesion-level endpoints, calibration, and decision-curve analysis to connect image quality with patient benefit. (v) Prospective trials: assess impact on referral decisions and treatment planning, compare against dye-based FA when ethically feasible, and evaluate operational benefits in teleophthalmology workflows. In summary, phase-aware synthesis of UWF_FA from UWF_RI yields high-fidelity, perceptually convincing angiograms and confers measurable gains in DR stratification when integrated with a transformer-based classifier. By reducing reliance on dye injection while preserving angiographic insight, the approach offers a scalable path to safer, more accessible screening. Continued work on domain robustness, lesion-aware validation, and prospective clinical studies will be key to realizing translational impact at scale.