Denoising swin transformer and perceptual peak signal-to-noise ratio for low-dose CT image denoising

Abstract

Low-dose CT examination will reduce the radiation risk to patients, but noise and artifacts in CT images will increase. This paper proposes a Denoising Swin Transformer (DnST) model for low-dose CT image denoising. DnST model modifies the network structure based on Swin Transformer and introduces a perceptual loss function and residual mapping into the network. In addition, this paper also designs a new image quality assessment (IQA) metric, PPSNR (Perceptual Peak Signal-to-Noise Ratio), which measures the difference of higher-level information of image and optimizes the problem that the scores of PSNR are inconsistent with the visual quality seen by human eyes. The experimental results show that compared with the CNN method, DnST achieves good results in noise suppression, structure protection and lesion detection. The image quality is improved by about 2.28% ∼15.95%. Not only noise is suppressed, but also the structure and critical information are protected.