DICDNet: A Convolutional Dictionary-based Metal Artifact Reduction Network for CT Images

This research explores the observation that metal artifacts in CT images consistently display non-local streaking and star-shape patterns. To address this, the authors propose a convolutional dictionary model to encode these artifacts. A novel optimization algorithm, based on the proximal gradient technique, is introduced to solve the model. The algorithm's design utilizes simple operators, facilitating easy unfolding into network modules with specific physical meanings. This feature enhances the algorithm's interpretability. The authors refer to their model as DICDNet (Convolutional Dictionary-based Metal Artifact Reduction Network). DICDNet's effectiveness is rigorously assessed through comprehensive experiments on both synthesized and clinical datasets. The results demonstrate DICDNet's superiority over existing state-of-the-art metal artifact reduction (MAR) methods. DICDNet also stands out for its superior interpretability, providing a deeper understanding of its performance and underlying mechanisms.