Abstract

Fundus image quality is crucial for diagnosing eye diseases, but real-worldconditions often result in blurred or unreadable images, increasing diagnosticuncertainty. To address these challenges, this study proposes RetinaRegen, ahybrid model for retinal image restoration that integrates a readabilityclassifi-cation model, a Diffusion Model, and a Variational Autoencoder (VAE).Ex-periments on the SynFundus-1M dataset show that the proposed method achievesa PSNR of 27.4521, an SSIM of 0.9556, and an LPIPS of 0.1911 for thereadability labels of the optic disc (RO) region. These results demonstratesuperior performance in restoring key regions, offering an effective solutionto enhance fundus image quality and support clinical diagnosis.

 

Quick Read (beta)