Annotation Guided Generative Adversarial Networks for Image Augmentation of Age-Related Macular Degeneration

Age-related macular degeneration (AMD) is a common eye disease that affects millions of people worldwide, leading to central vision loss. It mostly occurs in people over 50 years old. As per the literature, AMD diagnosis requires high-quality and diverse fundus images, which are often scarce and expensive to obtain. We propose a novel generative adversarial network (GAN) to address this challenge. It can synthesize realistic and clinically relevant images with AMD anomalies (drusen and Abnormalities of Choroidal Neovascular Membrane(CNVM)), enhancing dataset diversity and aiding automated diagnosis. Our GAN architecture is supplemented with (manual annotation guided) spatial attention maps to emphasize on anomalies guided through manually generated label maps. Further we propose 2 Siamese like Discriminators, for individual anomalies which emphasize the Generator module to generate AMD anomaly (drusen and abnormalities of CNVM) specific images. The proposed GAN leverages various loss functions, including contrastive loss and anomaly loss in the Siamese like Discriminator, and the adversarial loss in the combined model. The anomaly loss inspired by (Retinal detail Loss) measures the prominence of AMD signatures in the generated images. Our approach successfully generates synthetic images with specific AMD anomalies, aiding early diagnosis. The proposed GAN extended excellent dataset augmentation capabilities with 3% and 4% increase in diagnostic accuracy of drusen and abnormalities of CNVM, respectively.