A Dual-Domain Parallel Fusion Network for Low-Dose CT Reconstruction

Low-dose computed tomography (LDCT) imaging is of clinical significance to reduce X-ray radiation risk to human health. However, it brings noise and lowers the image quality, which is unfavorable to clinical diagnostics. In this paper, we present a new dual-domain parallel fusion network (DPFNet) to make high-quality LDCT reconstruction. It consists of three main parts: 1) a residual dense network with attention for sinogramdomain denoising; 2) a dynamic serpentine convolutional U-Net for image-domain feature extraction; 3) a dual-domain information fusion module to make joint mutual optimization with complementary information interaction between two domains. Extensive experiments on the publicly available NIH-AAMPMayo Clinic LDCT dataset demonstrate that the proposed DPFNet can achieve promising PSNR up to 44.89 dB, SSIM up to 98.20%, and MAE down to 9.58 HU, as well as high-quality visualizations for both quantitative and qualitative assessments.