Abstract
Deep learning methods provide state-of-the-art performance for the semantic segmentation of the retina and choroid in optical coherence tomography (OCT) images, enabling rapid, accurate and automatic analyses. However, high difficulty scans can still pose a problem even for the current state-of-the-art methods. Generative adversarial networks (GANs), are a family of deep learning methods that provide significant benefits for several applications due to their ability to learn complex data distributions, such as those of large image datasets. Segmentation is one of these applications that has been investigated in several modalities including retinal fundus image analysis, resulting in performance improvements when incorporating an adversarial loss for segmentation. However, the application of GAN-based segmentation to OCT images has not been investigated in detail and has not been studied at all in the context of choroidal segmentation. In this study, we investigate the use of a GAN to perform semantic segmentation of the retina and choroid in OCT images, by replacing the traditional segmentation loss with an adversarial loss. A detailed analysis of important training parameters and network architecture choices is provided to 1) better understand their behavior and 2) to optimize performance for chorio-retinal segmentation in OCT images. A key difference of this study is that, by considering the loss in isolation and comparing to traditional segmentation losses using an identical segmentation network, an unbiased and transparent comparison is performed. Using an optimized adversarial loss, strong performance is observed, providing near comparable performance to traditional segmentation losses. The results from this experiment provide a strong foundation for future work with GAN-based OCT retinal and choroidal segmentation.
