Enhancing OCT patch-based segmentation with improved GAN data augmentation and semi-supervised learning

Abstract

For optimum performance, deep learning methods, such as those applied for retinal and choroidal layer segmentation in optical coherence tomography (OCT) images, require sufficiently large and diverse labelled datasets for training. However, the acquisition and labelling of such data can be difficult or infeasible due to privacy reasons (particularly in the medical domain), accessing patient images such as those with specific pathologies, and the cost and time investment to annotate large volumes of data by clinical experts. Data augmentation is one solution to address this issue, either using simple variations and transformations of the images (e.g. flips, brightness) or using synthetic data from sophisticated generative methods such as generative adversarial networks (GANs). Semi-supervised learning (SSL) is another technique which aims to utilise unlabelled data to enhance the performance of deep learning methods and is beneficial where significant amounts of data may be available but are not labelled. In this study, we aim to enhance patch-based OCT retinal and choroidal layer segmentation with both GAN-based data augmentation and SSL. In particular, we employ a conditional StyleGAN2 to generate synthetic patches for data augmentation and a similar unconditional GAN for pre-training the patch classifier to perform SSL. In doing so, we propose a new patch classifier architecture based on the discriminator architecture to improve performance, in addition to the SSL benefit. Compared to previous methods, the proposed data augmentation approach provides an improved data augmentation performance for patch classification with its effectiveness widespread, particularly in the case of low data, across three different OCT datasets encompassing a range of scanning parameters, noise levels, pathology and participant variability. The method provides some subsequent improvements in boundary delineation which is of high importance from a clinical perspective. Additionally, the proposed SSL approach boosts classification performance and boundary delineation performance in some cases which provides further usefulness in the case of low data. The proposed methods can be utilised to enhance OCT segmentation methods, which may be of considerable benefit for both clinicians and researchers.