SPEAR-Net: Self-Prior Enhanced Artifact Removal Network for Limited-Angle DECT

Abstract

Dual-energy computed tomography (DECT) is a fully functional instrument for disease detection in clinical practice because of its ability to identify substances and quantify materials. In some practical applications, due to the limitation of scanning conditions, projection data can only be collected from a limited angle, and the consistency of measurement cannot be guaranteed. Existing dual-energy CT reconstruction methods fail to address well the severe artifacts and noise in dual-energy CT images caused by limited-angle scanning. In this paper, we proposed a Self-Prior Enhanced Artifact Removal Network (SPEAR-Net) for limited-angle DECT, which can effectively combine the complementary information in the high- and low-energy domains and self-prior information to contribute positively to the reconstruction of high-quality dual-energy CT images. The SPEAR-Net consists of an image-domain self-prior network(IP-Net), two dual-energy image-domain networks(DIP-Net), and a dual-energy sinogram-domain self-prior network(DSP-Net). Specifically, the IP-Net and DIP-Net are adopted to extract the features of the dual-energy CT reconstructed images under dual-quarter scanning as prior information. The self-prior projection obtained from the forward projection of the prior CT image is harnessed by DSP-Net to complete the dual-energy limited-angle projection data and to facilitate the performance of SPEAR-Net in removing artifacts in the reconstructed dual-energy images. Qualitative and quantitative analyses demonstrate the superior capability of SPEAR-Net in dual-energy limited-angle projection data complementation, detail preservation, and artifact removal. Two popular DECT applications, virtual non-contrast (VNC) imaging and iodine contrast agent quantification, reveal that images reconstructed by SPEAR-Net have promising clinical significance.