Generative Adversarial Networks with Dense Connection for Optical Coherence Tomography Images Denoising

Abstract

Optical coherence tomography (OCT) is widely used in the diagnosis of ophthalmic diseases. However, OCT is affected by ubiquitous speckle noise which make it difficult to analysis the retinal structures. To efficiently remove the noise as well as preserve clinical detail information contained in the images, we suggest to train a denoise generative adversarial network (DNGAN) jointly with a densely connected convolutional network to estimate clean OCT images from noisy OCT images. A generator convolutional neural network (CNN) with several dense connections, is trained to transform noisy OCT image into clean OCT image. At the same time, an adversarial CNN is trained to improve the denoising performance of the generator. The experimental results demonstrate the superior performance of our network on OCT images.