NLUNet: A Novel Approach for Optical Coherence Tomography Images Enhancement via Deep Learning

Abstract

Speckle noise significantly reduces the quality of optical coherence tomography (OCT) images. In this paper, two new methods, SRUNet and NLUNet, are introduced to enhance OCT images and reduce speckle noise. These methods are developed based on deep learning and a conventional image processing technique while operating without ground truth images. In the SRUNet method, a combination of UNet and ResNet architectures is used, so that the UNet encoder section is combined with ResNet to reduce speckle noise and recover structural details with significant accuracy. In the NLUNet method, the UNet architecture combined with a nonlocal mean filter (used as a post-processing) resulted in the removal of speckle noise and improved image resolution. These two methods were evaluated using PSNR, SSI and MSE metrics. The measured values of PSNR, SSIM and MSE were 28.88, 0.76 and 84.05, respectively. The results demonstrate that the NLUNet method performs properly to reduce the noise and improve the image quality.