Low-Dose CT Image Denoising Using Cycle Consistent Adversarial Networks

Abstract

Computed tomography (CT) has been widely used in 
modern medical diagnosis and treatment. However, 
ionizing radiation of CT for a large population of 
patients becomes a concern. Low-dose CT is actively 
pursued to reduce harmful radiation, but faces 
challenges of elevated noise in images. To address 
this problem and improve low-dose CT image 
quality, we develop an image-domain denoising 
method based on cycleconsistent adversarial 
networks (CycleGAN). Different from previous deep 
learning based denoising methods, CycleGAN can 
learn data distribution of organ structures from 
unpaired full-dose and low-dose images, i.e. there is 
no one-to-one correspondence between full-dose 
and low-dose images. This is an important 
development of learning-based methods for low
dose CT since it enables the model growth using 
previously acquired full-dose images and later 
acquired low-dose images from different patients. 
As a proof-of-concept study, we used the NIH
AAPM-Mayo Clinic Low Dose CT Grand Challenge 
data to test our CycleGAN denoising method. The 
results show that the proposed method not only 
achieves better peak signal-to-noise ratio (PSNR) 
for quarter-dose images than non-local mean and 
dictionary learning denoising methods, but also 
preserves more details reflected by images and 
structural similarity index (SSIM). Our investigation also reveals that a larger sample size leads to a 
better denoising performance for CycleGAN.