OCT retinal image-to-image translation : Analysing the use of CycleGAN to improve retinal boundary semantic segmentation

Optical coherence tomography (OCT) images of the posterior eye provide valuable clinical information. To quantify these images and extract appropriate biomarkers, methods to segment the different retinal boundaries are needed. In recent years, deep learning methods have been applied to perform this image analysis task, providing state of the art performance. However, these methods can be affected by image variability, particularly if the network is trained with images that do not match in terms of features to those of the testing dataset. One of the common sources of variability in OCT is speckle noise. In this work, the effect of noise on the semantic segmentation process is investigated and the use of a CycleGAN method as an image-to-image translation to reduce noise and its impact on segmentation are assessed. The results show promising performance and a proof of the potential of this generative adversarial network (GAN) method to positively impact medical image segmentation, obtaining good performance results when compared in terms of Dice coefficient overlap and boundary error metrics. The findings of this work may be translated to other applications such as 'OCT instrument translation' to create instrument-agnostic segmentation solutions.