Deep Learning Based Retinal OCT Image Denoising using Generative Adversarial Network

摘要

Optical Coherence Tomography (OCT) is the mostly used imaging modality for detecting retinal diseases. But, the presence of multiplicative granular type speckle noise in the OCT image makes it difficult to accurately identify the retinal disorders. This paper presents, a computer vision and deep learning based OCT image denoising technique using Generative Adversarial Network (GAN) that contains a generator and a discriminator. The generator generates denoised image form noisy input image and the discriminator acts as a classifier whether the image is real or denoised image. Furthermore, an improved loss function has been added to train the GAN network. Through this adversarial training process, weight of the generator gets updated and produces denoised image with much similar as ground-truth image. A UNet shaped Res-Autoencoder and a CNN model is used as generator and discriminator respectively. To evaluate the proposed system, the peak signal-to-noise ratio (PSNR) and mean-square error (MSE) were used as evaluation matrices. In addition to that, the denoising performance of the proposed model was compared with some traditional image denoising techniques such as Wavelet-transform, Bilateral, Non-local Means (NLM) and BM3D. The evaluation matrices validate the supremacy of the acclaimed model on the OCT image.