UWF-RI2FA: Generating Multi-frame Ultrawide-field Fluorescein Angiography from Ultrawide-field Retinal Imaging Improves Diabetic Retinopathy Stratification

Ultrawide-field fluorescein angiography (UWF-FA) facilitates diabetic retinopathy (DR) detection by providing a clear visualization of peripheral retinal lesions. However, the intravenous dye injection with potential risks hamper its application. We aim to acquire dye-free UWF-FA images from noninvasive UWF retinal imaging (UWF-RI) using generative artificial intelligence (GenAI) and evaluate its effectiveness in DR screening. A total of 18,321 UWF-FA images of different phases were registered with corresponding UWF-RI images and fed into a generative adversarial networks (GAN)-based model for training. The quality of generated UWF-FA images was evaluated through quantitative metrics and human evaluation. The DeepDRiD dataset was used to externally assess the contribution of generated UWF-FA images to DR classification, using area under the receiver operating characteristic curve (AUROC) as outcome metrics. The generated early, mid, and late phase UWF-FA images achieved high authenticity, with multi-scale similarity scores ranging from 0.70 to 0.91 and qualitative visual scores ranging from 1.64 to 1.98 (1=real UWF-FA quality). In fifty randomly selected images, 56% to 76% of the generated images were difficult to distinguish from real images in the Turing test. Moreover, adding these generated UWF-FA images for DR classification significantly increased the AUROC from 0.869 to 0.904 compared to the baseline model using UWF-RI images (P < .001). The model successfully generates realistic multi-frame UWF-FA images for enhancing DR stratification without intravenous dye injection.

💡 Research Summary

The study addresses a critical limitation of ultra‑wide‑field fluorescein angiography (UWF‑FA): the need for intravenous dye injection, which carries risks and limits routine use in diabetic retinopathy (DR) screening. Leveraging a large in‑house dataset of 1,352 patients, the authors paired 2,747 ultra‑wide‑field retinal images (UWF‑RI) with 18,321 corresponding FA images across early, mid, and late phases. After meticulous preprocessing—masking peripheral artifacts, extracting vascular maps with the RMHAS system, and performing pixel‑wise registration using AKAZE keypoints and RANSAC—they retained only high‑quality pairs (Dice ≥ 0.5).

A generative adversarial network based on pix2pixHD was adapted to translate UWF‑RI into phase‑specific FA images. To enhance high‑frequency details such as fine vessels and lesions, Gradient Variance Loss was added to the original adversarial and reconstruction losses. Three independent models (one per phase) were trained on 1,024 × 1,024‑pixel inputs with extensive augmentation (random crops, flips, rotations). Training used a batch size of 4, learning rate 2 × 10⁻⁴, and 50 epochs.

Quantitative metrics on an internal test set showed mean absolute error (MAE) of 100–114, peak signal‑to‑noise ratio (PSNR) of 27–29 dB, structural similarity index (SSIM) of 0.84, and multi‑scale SSIM of 0.88–0.91, indicating high fidelity to real FA. Human evaluation by two ophthalmologists on 50 image sets yielded average visual quality scores of 1.64–1.98 on a 1‑to‑5 scale (1 = indistinguishable from real), with Cohen’s κ ranging from 0.79 to 0.84, reflecting substantial inter‑rater agreement. In a Turing test, experts correctly identified real FA 80‑88% of the time, but misclassified generated FA as real in 56‑76% of cases, underscoring the realism of the synthetic images.

To assess clinical utility, the authors applied the model to the publicly available DeepDRiD dataset (256 UWF‑RI images with DR severity labels). Synthetic FA images for each phase were generated and combined with the original RI as inputs to a DR classification pipeline consisting of a Swin Transformer feature extractor and a multilayer perceptron classifier. Four experimental conditions were compared: (1) RI only (baseline AUROC 0.869), (2) RI + early FA (AUROC 0.886), (3) RI + early + mid FA (AUROC 0.887), and (4) RI + early + mid + late FA (AUROC 0.904). All additions significantly improved AUROC, area under the precision‑recall curve (AUPR), and F1 scores (P < 0.001). The progressive performance gains demonstrate that multi‑phase synthetic FA provides complementary vascular information beyond what is captured in color fundus images alone.

The paper’s contributions are threefold: (1) a robust cross‑modal registration pipeline for ultra‑wide‑field images, (2) a GAN‑based translation framework that can synthesize realistic, multi‑phase FA without dye injection, and (3) empirical evidence that these synthetic images enhance automated DR severity grading. Limitations include reliance on data from a single clinical site, modest external validation (only DeepDRiD), and potential residual gaps in reproducing subtle leakage or micro‑aneurysm patterns that are not evident in RI. Future work should involve multi‑center studies, broader device compatibility testing, and prospective clinical trials to determine whether synthetic FA can replace or augment real FA in routine DR screening workflows. Overall, the study showcases a promising, cost‑effective, and patient‑friendly avenue for improving diabetic retinopathy detection using generative AI.