Sparse-sampling CT Sinogram Completion using Generative Adversarial Networks

Sparse-sampling for low-dose Computed tomography (CT) is currently a subject of extensive investigations due to its potential to reduce the radiation dose. However, severe streak artifacts can be observed when reconstructing the sparse-sampled data with conventional FBP algorithms. Solutions to this problem fall into two categories - interpolation methods in the sinogram domain and iterative methods for reconstruction. The former often yield results with residual streak artifacts, and the latter require ample time and computer memory. In this work, we propose to use a deep learning method to complete the sparse-sampled sinogram. The network is named Pix2Pix, which is a conditional GAN structure. Results show our approach can accurately complete sinograms with excellent generalization ability. The synthesized sinograms can thus be reconstructed by FBP without streak artifacts.