Fundus Image Generation Based on Generative Adversarial Networks with Dual Discriminators

Generative adversarial networks for fundus image generation can address the scarcity of fundus images, reduce preparation costs, and protect patient privacy. Existing methods for generating adversarial networks often have drawbacks, such as unrealistic local details, unclear resolution, and reliance on conditional inputs like blood vessel annotations. To overcome these limitations, this paper proposes a generative adversarial network with dual discriminators that generate fundus images from random noise. This network consists of a generator and two distinct discriminators that consider both the global features and local details of fundus images. Extensive experiments validate the proposed method. The inception score, Fréchet Inception Distance, and Kernel Inception Distance achieved by our method are 1.882, 80.52, and 0.0681, respectively. These results indicate that our method can generate fundus images with diverse, realistic, and clear local details, as well as high overall quality.