Image-to-Image Translation Based CT Metal Artifact Reduction

As a non-invasive examination method,Computed Tomography(CT)is widely used in imaging diagnosis,radiotherapy planning,surgical navigation and other clinical tasks.However,during CT scan,if the scanned area contains metal materials,it will be affected by beam hardening,photon starving,and other physical effects,and the reconstructed CT image will be involved by light-dark radial stripes,namely metal artifacts.Metal artifacts can make the wrong CT values displayed around the metal object,blur the image details,make the boundary disjunctive line unclear,produce poor visual effect,reduce the image quality,affect the doctor’s diagnosis,and bring problems to the calculation of radiation dose.Therefore,it is of great clinical significance to explore the method of correcting metal artifacts in CT images.In order to address the problem of metal artifact reduction,this paper proposes two algorithms for metal artifact reduction based on image-to-image translation.To effective reduce the metal artifacts,the original anatomical structure information contained in CT images,especially the tissue structure information around metal objects,should be restored as far as possible to improve the quality of corrected images.Specifically,the research content is as follows:Firstly,an artifact correction model combining sinogram domain interpolation with unsupervised image-to-image translation(UNIT)network is proposed.This model is aimed to use unpaired prior images for secondary artifact correction caused by projection domain interpolation,preserve the details of the organization structure near the metal,and improve the quality of the image.Firstly,the linear interpolation method is used to correct the metal artifacts in the sinogram domain.Then,based on prior images,the UNIT network is formed by combining the generation adversarial network(GANs)and variational autoencoder(VAEs)to eliminate the secondary artifacts generated by the simple sinogram domain correction and restore the details of the image.In this paper,the preliminarily corrected image and the prior image without artifact are regarded as elements in two different image domains,and the artifact removed image can be obtained through the image translation network.In order to verify the effectiveness of the proposed method,simulated data are used for experiments.PSNR and SSIM were used as quantitative indicators to evaluate the correction results.Compared with LI algorithm,the PSNR and SSIM of the proposed algorithm are improved by 13.7566 and 0.0557 when the metal is small.When the metal is large,the PSNR and SSIM of the proposed algorithm are improved by 9.2794 and 0.1783.The method proposed in this paper achieves better results and retains more details of the image organization structure compared with other methods of artifact removal.Secondly,this paper presents a metal artifact removal method for prior images based on Mega-voltage Computed Tomography(MVCT)image transformation.This paper aims to use MVCT images of the same patient to recover the image details damaged by metal artifacts in CT images.Specifically,the MVCT image is converted to sKVCT image by image translation network first.The sKVCT image is applied to the NMAR algorithm as a prior image for metal artifact reduction.Then the corrected image is put into the U-net network to enhance image detail and improve image quality.In this paper,simulate data of phantom and patients is used for experiments,and PSNR and SSIM are selected as objective evaluation indexes.Compared with LI algorithm,the PSNR and SSIM of the proposed algorithm are improved by 16.5274 and 0.2213when the metal is small.When the metal is large,the PSNR and SSIM of the proposed algorithm are improved by 15.2109 and 0.2158.Experiments show that the method proposed in this paper has the best effect in correcting metal artifacts,with clear edges and details.