Investigation of Low-Dose CT Image Denoising Using Unpaired Deep Learning Methods

Low-dose computed tomography (LDCT) is desired due to prevalence and ionizing radiation of CT, but suffers elevated noise. To improve LDCT image quality, an image-domain denoising method based on cycle-consistent generative adversarial network (CycleGAN) is developed and compared with two other variants, IdentityGAN and GAN-CIRCLE. Different from supervised deep learning methods, these unpaired methods can effectively learn image translation from the low-dose domain to the full-dose (FD) domain without the need of aligning FDCT and LDCT images. The results on real and synthetic patient CT data show that these methods can achieve peak signal-to-noise ratio (PSNR) and structural similarity index (SSIM) comparable to, if not better than, the other state-of-the-art denoising methods. Among CycleGAN, IdentityGAN, and GAN-CIRCLE, the later achieves the best denoising performance with the shortest computation time. Subsequently, GAN-CIRCLE is used to demonstrate that the increasing number of training patches and of training patients can improve denoising performance. Finally, two nonoverlapping experiments, i.e., no counterparts of FDCT and LDCT images in the training data, further demonstrate the effectiveness of unpaired learning methods. This work paves the way for applying unpaired deep learning methods to enhance LDCT images without requiring aligned FD and low-dose images from the same patient.

Automated summary

The study proposes an image-domain denoising method based on CycleGAN that can effectively learn image translation from low-dose domain to full-dose domain without aligning FDCT and LDCT images. Experimental results show that this method achieves good denoising performance in LDCT image processing, comparable to or better than other state-of-the-art denoising methods.