Assessment of Deep Learning-Generated Ultra-Widefield Fluorescein Angiography From Fundus Images in Diabetic Retinopathy

Purpose: To evaluate and compare the clinical validity of synthetic ultra-widefield fluorescein angiography (UWFA) images generated from ultra-widefield fundus photography (UWFP) using generative adversarial networks (GANs) in patients with diabetic retinopathy (DR). Methods: Two GAN-based models, RegGAN and UWAFA-GAN, were trained to generate synthetic UWFA images from corresponding UWFP acquired using Optos California P200DTx. The dataset included 2084 image pairs (no DR: 124; nonproliferative DR: 795; severe NPDR: 770; proliferative DR: 395). Technical image similarity was assessed using peak signal-to-noise ratio (PSNR) and structural similarity index (SSIM). Clinical similarity was assessed by expert graders using the DR severity scale, comparing the synthetic images with corresponding real UWFA images. Results: Both models demonstrated acceptable performance in generating synthetic UWFA images from UWFP (F1-score range: 0.281-0.709; SSIM: 0.43-0.585; PSNR: 18.15-19.90). UWAFA-GAN achieved higher quantitative similarity metrics (SSIM and PSNR), indicating superior overall image fidelity, whereas RegGAN showed stronger clinical correlation with ground-truth UWFA, achieving higher diagnostic accuracy (66.7% vs. 49.2%) and more balanced precision, recall, and F1-scores across DR severity levels. This suggests that while UWAFA-GAN excels in pixel-level resemblance, RegGAN better preserves diagnostically relevant features. Conclusions: GAN models successfully generated UWFA images from UWFP in patients with DR. With further improvements, such models may serve as a complementary noninvasive imaging tool in selected clinical settings and may extend accessibility in resource-limiting settings. Translational Relevance: GAN-based synthetic angiography may serve as a noninvasive alternative to UWFA, particularly when conventional angiography is contraindicated or unavailable.