CD-Net: Comprehensive Domain Network With Spectral Complementary for DECT Sparse-View Reconstruction

Abstract

Dual-energy computed tomography (DECT) is of great clinical significance because of its material identification and quantification capacity. Although DECT measures attenuation using two different spectra, the anatomical structure of the low- and high-energy CT images are consistent with each other and the images are also correlated in the energy domain, resulting in significant information redundancy. Here this redundancy has been exploited by the proposed CD-Net (Comprehensive Domain Network) with spectral complementarity to improve the image quality and reduce the radiation dose for DECT imaging. CD-Net restores accurate anatomical information from both projection and image domains. It encompasses a projection domain neural network (PD-Net), an analytical reconstruction operator (ARO) and an image domain neural network (ID-Net). Embedding ARO into deep learning framework, the proposed one-step DECT sparse reconstruction method can directly produce high-quality DECT images from projection data acquired with spectral complementarity scheme. Qualitative and quantitative analyses demonstrate the competitive performance of CD-Net in terms of CT number accuracy, detail preservation and artifact removal. Two popular DECT applications, virtual non-contrast (VNC) imaging and iodine contrast agent quantification, reveal that the images reconstructed by CD-Net are promising for clinical applications.