Deep Learning-based Prior toward Normalized Metal Artifact Reduction in Computed Tomography

X-ray computed tomography (CT) often suffers from scatter and beam-hardening artifacts in the presence of metal. These metal artifacts are problematic as severe distortions in the CT images deteriorate the diagnostic quality in clinical applications such as orthopedic arthroplasty. The normalized metal artifact reduction (NMAR) method effectively reduces the artifacts by normalizing the sinogram with the metal traces through the forward projection of the prior image. Because the prior image is the thresholded CT image with the values of the air and soft tissues replaced, the image is noticeably different from the ideal CT thereby making normalized sinogram not completely flat. In this paper, we propose the novel NMAR method with the deep learning-enhanced prior image which is denoised by learning the relationship between NMAR and clean image without metal artifact. The denoised prior image is then forward projected to correct the sinogram with the metal trace. The experimental results on simulated rat phantom dataset demonstrate that our proposed deep prior NMAR achieves higher structural similarity index (SSIM) and peak signal-to-noise ratio (PSNR) than the original NMAR.