Dual image and mask synthesis with GANs for semantic segmentation in optical coherence tomography

In recent years, deep learning-based OCT segmentation methods have addressed many of the limitations of traditional segmentation approaches and are capable of performing rapid, consistent and accurate segmentation of the chorio-retinal layers. However, robust deep learning methods require a sufficiently large and diverse dataset for training, which is not always feasible in many biomedical applications. Generative adversarial networks (GANs) have demonstrated the capability of producing realistic and diverse high-resolution images for a range of modalities and datasets, including for data augmentation, a powerful application of GAN methods. In this study we propose the use of a StyleGAN inspired approach to generate chorio-retinal optical coherence tomography (OCT) images with a high degree of realism and diversity. We utilize the method to synthesize image and segmentation mask pairs that can be used to train a deep learning semantic segmentation method for subsequent boundary delineation of three chorioretinal layer boundaries. By pursuing a dual output solution rather than a mask-to-image translation solution, we remove an unnecessary constraint on the generated images and enable the synthesis of new unseen area mask labels. The results are encouraging with near comparable performance observed when training using purely synthetic data, compared to the real data. Moreover, training using a combination of real and synthetic data results in zero measurable performance loss, further demonstrating the reliability of this technique and feasibility for data augmentation in future work.