DSEGAN: Detail and structure enhanced generative adversarial network for fundus image enhancement

Fundus imaging is an important tool for screening and diagnosing ophthalmic diseases. Despite continuous advancements in imaging technology, low-quality images may still occur due to environmental factors or patient conditions, which can adversely affect clinical diagnosability and downstream analysis. In clinical practice, it is difficult to simultaneously acquire paired high-quality and low-quality fundus images of the same patient, making supervised learning-based image enhancement methods difficult to apply. Existing unsupervised methods, such as CycleGAN, can achieve automatic mapping from low-quality to high-quality image domains without requiring paired images; however, they still have limitations in modeling the characteristics of fundus images, primarily manifested as the loss of vessel texture details and the blurring of optic disc contour structures. To address these challenges, we propose a Detail-Structure Enhanced Generative Adversarial Network (DSEGAN) for fundus image enhancement, which employs our designed Detail-Structure Enhanced Generator (DSEGen) as its core. In the encoding stage, DSEGen introduces an attention Enhanced Partial Convolution (AEPConv) block to suppress noise and highlight high-frequency information such as vessel texture details. In the decoding stage, a dual-Branch Fusion Block (DBFB) is adopted to achieve the collaborative fusion of high-frequency information (e.g., vessel details) and low-frequency information (e.g., optic disc contour structures). Experiments conducted on three real-world datasets (CFP, UWF, and EyeQ) and one synthetic dataset (FIVES) demonstrate that DSEGAN outperforms existing methods in objective metrics such as NIQE, FID, and KID. Moreover, its effectiveness in fundus image enhancement is further validated through a downstream vessel segmentation task on the synthetic dataset.