Retinal Fundus Image Synthesis using Image Segmentation and DCGAN

Training high-quality deep learning models in the ophthalmic domain requires many good-quality labeled training images. Unlike natural images, which are readily available, there is a scarcity of good quality labeled medical images. This work, address the problem for the ophthalmic domain by proposing a methodology to generate good-quality synthetic fundus images using available real images. This work utilizes image segmentation to separate important components of a fundus image, such as the vessel network and optic disc, then combine them with the extracted background to form low-quality fundus images. These images, along with the real image set, are used to train a Deep Convolutional Generative Adversarial Network (DCGAN). The resulting GAN model will be capable of generating high-quality fundus images. The synthetic images generated are of good quality and, at the same time, different from the training set used initially. These images could be employed in tandem with both the preliminary synthetic image and the authentic image to train a classification model. Moreover, they can provide a foundation for transfer learning in the classification of eye diseases.