Contrastive Learning for Generating Optical Coherence Tomography Images of the Retina

Abstract

As a self-supervised learning technique, contrastive learning is an effective way to learn rich and discriminative representations from data. In this study, we propose a variational autoencoder (VAE) based approach to apply contrastive learning for the generation of optical coherence tomography (OCT) images of the retina. The approach first learns embedding representation from data by contrastive learning. Secondly, the learnt embeddings are used to synthesize disease-specific OCT images using VAEs. Our results reveal that the diseases are separated well in the embedding space and the proposed approach is able to generate high-quality images with fine-grained spatial details. The source code of the experiments in this paper can be found on Github (https://github.com/kaplansinan/OCTRetImageGen_CLcVAE).