Beam hardening artifact correction in x-ray computed tomography of homogeneous objects using Residual U-Net

Abstract

Correction of beam hardening artifacts in the x-ray computed tomography of homogeneous objects was performed using a trained Residual U-Net instead of the usual linearization technique. The Residual U-Net was trained using simulated x-ray CT data of ninety homogeneous digital phantoms. The ninety phantoms were cast into fifteen different materials to allow the network to learn material non-specificity. Data augmentation was performed by breaking the simulated x-ray CT data into smaller patches. After training, the network was tested on a new phantom and the results were compared against the monochromatic and polychromatic reconstructions. Using the monochromatic reconstruction as the ground truth image, the global SSIM improved from 0.7866 in the polychromatic reconstruction to 0.9205 in the reconstruction from the data corrected by the trained network. The local SSIM maps also show the effect of correction wherein the regions of high difference were noticeably reduced. Further test cases have to be made to assess the capability of the trained network to accommodate other materials and object geometries.