IRDNet: Iterative Relation-Based Dual-Domain Network via Metal Artifact Feature Guidance for CT Metal Artifact Reduction

Abstract

The metal artifacts in computed tomography (CT) images not only affect diagnosis and treatment but also present a classic nonlinear inverse problem in CT reconstruction. In this study, we propose an iterative relation-based dual-domain network (IRDNet) that utilizes metal artifact feature guidance to reduce such artifacts in CT images. To the best of our knowledge, IRDNet leverages metal artifact features as guidance of the dual-domain network for the first time to reduce metal artifacts. Our framework incorporates artifact-corrupted and precorrected images (linear-interpolated images) as well as metal artifact features to effectively reduce metal artifacts for a high-quality prior CT image and corresponding prior sinogram. The prior image and prior sinogram are then iteratively recovered sinogram using the residual learning strategy and mitigate the artifacts of CT image with a metal-location guidance framework. We construct IRDNet in an unrolling manner to accurately optimize anatomical structures. Compared to the state-of-the-art algorithms, IRDNet consistently produces reasonable CT images with reduced metal artifacts, as evaluated both quantitatively and qualitatively across different-sized metal implant samples and different metal materials. It generalized different artifacts caused by metals of various sizes and materials and successfully recovered surrounding tissues. The experimental results demonstrate the potential of incorporating metal inherent features as priors in the dual-domain network for reducing metal artifacts.