Effect of Altered OCT Image Quality on Deep Learning Boundary Segmentation

Deep learning methods provide a platform to segment boundaries within the retina and choroid in OCT images of the posterior eye, with the ultimate goal of having a robust model that works well across a wide range of different datasets. However, since most studies of deep learning methods use datasets exhibiting similar image quality for both training and evaluation, the effect of varied image quality on such methods is not normally explored in the context of OCT image segmentation. An understanding of the effects of image quality factors is vital to determine the robustness of the methods and their ability to be applied in clinical practice where images exhibiting a range of different qualities are encountered. This study examined a range of factors that can affect standard OCT image quality and determined how and why the performance of an existing neural network based segmentation method can subsequently degrade as a result. Three image quality factors (noise, contrast reduction, and gamma correction) all had a negative impact upon performance, while more robust performance was maintained in the presence of both JPEG and JPEG2000 image compression. Improving the method's robustness to each of these degradations is also demonstrated with marked performance improvements identified by applying a fine-tuning approach to the network. This study improves our understanding of the effect of OCT image degradation on neural network performance, the effect that fine-tuning with poor-image quality data has on the network and highlights the benefit and importance of training resilient models using data augmentation.