Noise reduction of fingertip OCT image based on generated unpaired high-quality datasets

Abstract

Optical Coherence Tomography (OCT) has been widely applied for noninvasive cross-sectional imaging in biological systems. When the incident light of OCT system meets with the particles in the internal tissues of organisms, light interference inevitably occurs, and irregular granular speckle noise will be formed on the generated scanning image. The speckle noise will cover the important subcutaneous tissue structure, so it is necessary to denoise the OCT image to reduce data loss. However, it is difficult to obtain high-quality clean samples of OCT images, so the traditional neural network denoising method has not achieved good results in this problem. This paper proposes a noise reduction method of fingertip OCT image based on generated unpaired high-quality datasets. In order to solve the problem of lacking noiseless images as truth values for network training, firstly, a small number of high-quality noiseless images are obtained by artificial enhancement. Then SinGAN network is used to train a single noiseless image to generate a large number of images with similar styles and different structures as truth values and create a dataset. Finally, CycleGAN is used to transform OCT images from noise domain to clean domain on unpaired datasets. Experiments show that this method can effectively remove speckle noise in OCT images and is superior to the traditional neural network method.