Constructing synthetic chorio-retinal patches using generative adversarial networks

The segmentation of tissue layers in optical coherence tomography (OCT) images of the internal lining of the eye (the retina and choroid) is commonly performed for clinical and research purposes. However, manual segmentation of the numerous scans is time consuming, tedious and error-prone. Fortunately, machine learning-based automated approaches for image segmentation tasks are becoming more common. However, poor performance of these methods can result from a lack of quantity or diversity in the data used to train the models. Recently, generative adversarial networks (GANs) have demonstrated the ability to generate synthetic images, which may be useful for data augmentation purposes. Here, we propose the application of GANs to construct chorio-retinal patches from OCT images which may be used to augment data for a patch-based approach to boundary segmentation. Given the complexity of GAN training, a range of experiments are performed to optimize performance. We show that it is feasible to generate 32×32 versions of such patches that are visually indistinguishable from their real variants. In the best case, the segmentation performance utilizing solely synthetic data to train the model is nearly comparable to real data on all three layer boundaries of interest. The difference in mean absolute error for the inner boundary of the inner limiting membrane (ILM) [0.50 vs. 0.48 pixels], outer boundary of the retinal pigment epithelium (RPE) [0.48 vs. 0.44 pixels] and choroid-scleral interface (CSI) [4.42 vs. 4.00 pixels] shows the performance using synthetic data to be only marginally inferior. These findings highlight the potential use of GANs for data augmentation in future work with chorio-retinal OCT images.