Synthesizing Realistic ARMD Fundus Images Using Generative Adversarial Networks (GANs)

Abstract

Age-related macular degeneration (ARMD) is an eye disease that can result in blurred or no vision in the central vision. It happens when aging causes damage to the macula and is the main reason for loss of sight for older adults. Deep learning techniques are widely used in ophthalmology, such as diagnosing age-related macular degeneration (ARMD), which requires a huge image dataset. But the existing datasets are unclear and insufficient for building the training models and require more preprocessing time. Also, Indian datasets are not used prominently for the leading causes of blindness and eye diseases. Therefore, this paper emphasizes on synthesizing large new datasets of artificial retinal images from the existing datasets by a deep learning approach generative adversarial networks which are the GANs. Generative adversarial networks (GANs) will be trained with fundus images from the age-related eye disease study (AREDS), producing synthetic fundus images with the ARMD. The performance of ARMD diagnostic DCNNs will be trained on the combination of both real and synthetic datasets. Images obtained by using generative adversarial networks (GANs) appear to be realistic and also increase the precision of the model. The deep learning model’s performance which uses the synthesized dataset should be close to the real images, suggesting that the dataset can be utilized for training humans and machines.