Deep learning has been successful in achieving good results for the metal artifact reduction (MAR) task in computed tomography (CT) images. However, most current deep learning frameworks lack sufficient model interpretability. Additionally, existing MAR techniques largely ignore the intrinsic a prior knowledge of metal artifact CT images, which could help to improve the effectiveness of MAR. This chapter presents an end-to-end convolutional dictionary super-resolution network for the MAR task. The network embeds the intrinsic a prior structure of metal artifacts and provides clear interpretability for the MAR task. Additionally, the generative adversarial structure of the network performs super-resolution reconstruction of the image with artifact removal to improve image clarity. Comprehensive experiments have confirmed the superior performance of the proposed network. It provides better metal artifact removal for CT images of different sites (pelvic, lungs, abdomen, and heart) and significantly improves the image quality.