An improved dual-domain network for metal artifact reduction in CT images using aggregated contextual transformations

Abstract

Objective. Metal artifact reduction (MAR) remains a challenging task due to the difficulty of removing artifacts while preserving anatomical details of the tissue. Although current dual-domain networks have shown promising performance in MAR, they heavily rely on the image domain, which can be too smooth and lose important information in the metal-affected area. To address this problem, we propose an improved dual domain network framework. Approach. We enhance sinogram completion performance by utilizing an aggregated contextual transformations network in the sinogram domain. Furthermore, we utilize a prior-projection-based linearized correction method to obtain images with beam-hardening artifacts removed, which are incorporated into the input of the image post-processing network to assist in training the image domain network. Finally, we train the sinogram domain network and the image domain network separately to their respective convergences. Main results. In experiments conducted on a simulated dataset, our method achieves the best average RMSE of 25.1, SSIM of 0.973, and PSNR of 42.1, respectively. Significance. The proposed method is capable of preserving tissue structures near metallic objects while eliminating metal artifacts from the reconstructed images. Related codes will be released at https://github.com/Corinna-China/AOTDudoNet