Voxel-sensitive Wavelet-based Approach for Structural Distortion in Low-Dose CT Images

Abstract

The decimation of high-frequency (HF) features during exhaustive low-dose computed tomography(LDCT) denoising introduces structural deformation. This paper addresses the aforementioned issue by introducing a GAN framework that provides novel adversarial training via discriminators in the wavelet and spatial domains. The wavelet-domain discriminator forces the generator to gain knowledge of the HF features via HF wavelet details (LH, HL, HH) and minimizes structural distortion. The generator network uses a spatial-domain discriminator to preserve local and global pixel correlations without altering the low-frequency (LF) features. Furthermore, we develop a generator network using a novel stationary wavelet-based residual block (SWTRB), which adaptively integrates spatial and frequency-domain information. In addition, we propose a wavelet-domain objective function on HF components, further improving the diagnostic quality of CT images. The experimental results demonstrate that the proposed method outperforms several state-of-the-art techniques on publicly available datasets, including "2016 NIH-AAPM-Mayo Clinic LDCT " and "Low-dose CT image and projection."