we first explore that the metal artifacts always presentnon-local streaking and star-shape patterns in CT imagesBased on such observations a convolutional dictionarymodel is deployed to encode the met

In this research, the authors start by noting that metal artifacts in CT images always exhibit non-local streaking and star-shape patterns. They propose a convolutional dictionary model to encode these artifacts. To solve the model, they introduce a new optimization algorithm based on the proximal gradient technique.

The proposed algorithm is designed to be easily unfolded into network modules with specific physical meanings, using simple operators. This allows for easier interpretation of the algorithm's steps. The authors refer to their model as DICDNet (Convolutional Dictionary-based Metal Artifact Reduction Network).

The effectiveness of DICDNet is evaluated through comprehensive experiments on both synthesized and clinical datasets. The results demonstrate the superiority of DICDNet compared to current state-of-the-art methods for metal artifact reduction (MAR). Additionally, DICDNet is shown to have superior interpretability, allowing for a better understanding of its performance and mechanisms