Deep Learning Super-Resolution Dictionary Network Removes CT Metal Artifacts

Deep learning has demonstrated efficacy in the task of Metal Artifact Reduction (MAR) within Computed Tomography (CT) images. Despite these successes, prevailing deep learning frameworks often fall short in terms of model interpretability. Moreover, existing MAR methodologies have largely overlooked the intrinsic prior knowledge within CT images that contain metal artifacts, knowledge which could significantly enhance the effectiveness of MAR. In response to these challenges, this paper introduces an end-to-end Super-resolution Convolutional Dictionary network tailored for MAR tasks. The network integrates the inherent structural priors of metal artifacts, thereby offering enhanced interpretability for the MAR process. Additionally, the network leverages a generative adversarial structure to perform super-resolution reconstruction on images, effectively mitigating artifacts and improving image clarity. Comprehensive experimental results substantiate the superior performance of the proposed network, which exhibits robust MAR capabilities across various anatomical regions (liver, pelvis, chest, abdomen) and markedly elevates image quality.