A low-light fundus image enhancement method based on GAN

Low-light enhancement in natural images has yielded a lot of results, especially in the presence of reference images. Unlike natural image enhancement, fundus image enhancement is more necessary to keep medical information such as fundus blood vessels unchanged in order to ensure subsequent and accurate medical diagnosis. However, direct application of natural image enhancement algorithms to the fundus image may result in the destruction of the fundus structure. And in clinical applications, there is a general lack of paired high and low-quality fundus images, which poses a further challenge to the study of augmentation methods. To address these problems, an efficient low-light fundus image enhancement model called BN-GAN was proposed. A low-light fundus image dataset was first established. Then, in order to solve the unpredictable brightness degradation, we propose the brightness normalization module. Furthermore, an unsupervised low-light fundus image enhancement method by fusing the attention mechanism in CycleGAN was designed. Finally extensive experiments demonstrate the superior effectiveness of our proposed method.