Compressive Spectral X-Ray CT Reconstruction via Deep Learning

Compressive spectral X-ray imaging (CSXI) uses coded illumination projections to reconstruct tomographic images at multiple energy bins. Different K-edge materials are used to modulate the spectrum of the X-ray source at various angles, thereby capturing the projections containing spectral attenuation information. It is a cost-effective and low-dose sensing approach; however, the image reconstruction is a nonlinear and ill-posed problem. Current methods of solving the inverse problem are computationally expensive and require extensive iterations. This paper proposes a deep learning model consisting of a set of convolutional neural networks to reconstruct the CSXI spectral images, which correspond to inpainting the subsampled sinograms, recovering the monoenergetic sinograms, and removing the artifacts from a fast but low-quality analytical reconstruction. Numerical experiments show that the proposed method significantly improves the quality of reconstructed image compared with that attained by the state-of-the-art reconstruction methods. Moreover, it significantly reduces the time-required for CSXI reconstruction.