A Feature-Guided Dual-Domain Network for CT Metal Artifact Reduction

Metal artifact reduction (MAR) in CT is a classical problem. Recently, deep learning-based (DL) methods have achieved promising results for MAR tasks, especially those methods based on dual-domain networks. However, current dual-domain networks seldom employ the inherent features of metal artifact during MAR, leaving room for further improvements. Here, we propose a dual-domain network guided by inherent metal artifact characteristics for MAR. Our method utilizes a feature extractor to obtain several artifact priors. Then, those priors were employed to guide a prior network to generate an accurate prior sinogram, which can subsequently guide the following dual-domain network. Our experimental results demonstrate the promise of the artifact feature-guided dual-domain network for the MAR in CT.