Supervised and Unsupervised Deep Learning Methods for Low-Dose CT Image Denoising

Low-dose Computed tomography (CT) is beneficial to patients if its diagnostic performance is not compromised. To suppress the elevated noise in low-dose CT images, deep learning based denoising methods have been actively pursued. In this study, we compared a supervised deep learning method, residual encoder-decoder convolutional neural network (RED-CNN), with unsupervised deep learning methods based on cycle-consistent generative adversarial networks (GAN), namely CycleGAN, GAN-CIRCLE and IdentityGAN. The unsupervised methods have the advantage of no need of the one-to-one correspondence between full-dose and low-dose training images over the supervised methods. We also investigate the influence of training samples on the denoising performance of two types of methods. The NIH-AAPM-Mayo Clinic Low Dose CT Grand Challenge data was used in this study. The results show that 1) the unsupervised methods can achieve the similar denoising performance as the supervised method for this data set, measured by peak signal-to-noise ratio (PSNR) and structural similarity index (SSIM); 2) increasing the sample size can improve the denoising performance in general; and 3) the simplest network structures in GAN-CIRCLE work best for the current denoising task in terms of PSRN and SSIM and speed.