Ring artifact removal for differential phase-contrast X-ray computed tomography using a conditional generative adversarial network.

Abstract

The integration process used as a pre-processing step in the reconstruction of differential phase-contrast X-ray CT (d-PCCT) causes the measurement noise to propagate throughout the projection image, which is leading to increased ring artifacts (RA) in the reconstructed image. It is difficult to eliminate the RA using conventional RA removal methods that were developed for the absorption-based CT field. We propose an effective method that can remove RA of d-PCCT images. The proposed method uses Laplacian images reconstructed from second-derivative projections of d-PCCT. This method is based on a conditional generative adversarial network (cGAN), whose loss function is designed by adding the L1- and L2-norm to the original cGAN. The training data were taken from a numerical phantom generated by a d-PCCT imaging simulator. To validate the applicability of the trained network, we tested its RA removal effect on test data from numerical phantoms generated randomly and actual experimental data. The results of numerical validation using numerical phantoms showed that the proposed method improved the RA removal effect compared to conventional methods. In addition, image comparison by visual evaluation showed that only the proposed method was able to remove RA while preserving original structures in the actual biological d-PCCT images. We proposed a cGAN-based method for RA removal that exploits the physical properties of d-PCCT. The proposed method was able to completely remove RA from d-PCCT images on both simulated data and biological data. We believe that this method is useful for the observation of various types of biological soft tissue.