Abstract
<title>Abstract</title> <p>Conversional fundus images characterized by blurred vessel boundaries, weak local textures, and poor visibility of fine structures can compromise the extraction of clinically relevant features for fundus disease diagnosing. To address this issue, we developed a super‑resolution enhancement method based on a generative adversarial network (SRGAN) augmented with an efficient channel attention (ECA) mechanism. The ECA modules, integrated into the residual blocks of the generator, selectively enhance feature responses around vessel edges, micro‑textures, and layered structures. To improve handle real‑world blur, we constructed a composite degradation model combining mild blur and bicubic down-sampling. Using a public diabetic retinopathy fundus dataset, we performed paired reconstruction experiments evaluated with PSNR, SSIM, and LPIPS for pixel fidelity, structural consistency, and perceptual quality. Additionally, on self‑collected clinical blurred fundus images, we applied no‑reference frequency‑domain and quality assessments including FRC Resolution, NIQE, and BRISQUE. Our method consistently outperformed baseline approaches in both synthetic and real‑world tests, improving the visibility of fine anatomical structures and overall image quality. These results suggest that SRGAN‑ECA with composite degradation and frequency‑domain evaluation offers a practical pathway for enhancing fundus image quality in clinical setting.</p>