A deep learning-based dual-domain information method for CT metal artifact reduction

When metal is present in the field of view of a CT scan, the reconstruction of images inevitably produces metal artifacts, significantly impacting image quality. In order to suppress metal artifacts, we propose a new deep learning CT metal artifact reduction (MAR) method that combines dual domain information from both the sinogram and image domains. Firstly, the adaptive optimal threshold segmentation method is used to segment the metal in the CT image and remove the metal corrosion area in the sinogram. Linear interpolation (LI) is used to preliminarily repair the missing metal area. After the metal-contaminated sinogram domain has been repaired using the sino-inpainting network, further picture information is recovered by employing an encoder-decoder network structure. The sinogram domain output from the network undergoes filtered back projection (FBP) to generate CT reconstructed images. To address inconsistencies in the initially corrected sinogram domain information, a non-local refine network is utilized in the image domain to reduce secondary artifact generation. This technique successfully lowers metal artifacts while maintaining image details, greatly improving the quality of the reconstructed images, according to experimental results using both simulated and real data.