Research On CT Image Metal Artifact Reduction Algorithm Based On Deep Learning

Computed tomography(Computed Tomography,CT)has a fast scanning speed and high imaging resolution.It can obtain internal three-dimensional structure information without damaging the detected object.It is widely used in the fields of medical clinical diagnosis,industrial non-destructive testing,and reverse design.However,due to beam hardening and Compton scattering,the metal inside the detected object will cause high-order nonlinear distortion of the attenuation information,resulting in metal artifacts in the reconstructed image.Metal artifacts appear as radial streak artifacts and black banded artifacts in CT images,which seriously affect the analysis and evaluation of images.The physical model for the accurate characterization of metal artifacts is complex and difficult to solve.The traditional correction method only uses the simplified physical model and a single set of data features.There are many deficiencies in secondary artifact suppression and metal edge information repair.Therefore,it is of great theoretical significance and practical value to study how to suppress metal artifacts and improve CT image quality.This paper studies metal artifact reduction methods that incorporate deep learning mechanisms.It uses data-driven methods to fully mined the intrinsic features of image domain artifacts and the distortion law of projection domain information.A series of theoretical,simulation and experimental studies are carried out focusing on the direct suppression of metal artifacts in the image domain,the complete correction of metal trace information in the projection domain,and the joint optimization correction of image and projection dual-domain of multi-metal complex artifacts.The main innovations obtained are as follows:1.A U-net based metal artifact reduction(U-MAR)in the image domain is proposed.Aiming at the problem that metal artifacts is difficult to be accurately described by fixed indicators or features,this paper designs an image domain metal artifact reduction network under the U-net architecture.Using multi-core convolution operation,the extraction and differentiation of various features in the image are realized by feature mining of metal artifact and tissue structure information in the image;the recovery of image tissue structure information is completed by using the decoding ability of the network based on feature information,and the correction of the metal artifact is realized by using the nonlinear fitting ability of the network through loss constraint;a large number of sample data containing artifacts are obtained by using the sample generation module based on the nonlinear metal artifacts simulation model,and univariate experimental analysis for network depth and convolution kernel size is carried out using the simulation dataset to further realize the optimization of network parameters.The experimental results show that the coding and decoding ability of the network makes the network suitable for image samples of different sizes.Compared with the normalized metal artifact reduction algorithm,the error in the correction results of this method is reduced by more than 10%.2.Sinogram Inpainting Metal Artifact Reduction(SI MAR)based on the completion network is proposed.In this paper,a projection-complement network is designed under the framework of deep convolution network by deeply mining the sinusoidal distribution characteristics of projection domain data and the nonlinear distortion law of metal projection.Using deep convolution and feature fusion,the complementary repair of distorted projections is achieved by fully extracting the global features of projection data and the distribution characteristics of metal projections;the consistency and continuity of the complementary data are further enhanced by designing the loss of image amplitude variance and the loss of projection sinusoidal characteristics.Combining projection characteristics and metal trace characteristics,using metal traces to zero,a medical simulation data set and a head model data set for network training and testing were made.The experimental results show that the mean absolute error of the projection completion results of this method is reduced by more than 20% compared with the U-net method,which can effectively complete the complete repair of metal traces,and the metal artifacts in the reconstructed images after the completion projection are significantly removed.3.A dual-domain Metal Artifact Reduction(Dual-D MAR)method based on generative adversarial networks is proposed.When the scanned object contains multiple metal implants,the projection data information will be seriously lost,resulting in increased completion error of the projection domain;at the same time,the image domain is more complex,resulting in reduced image domain correction performance.The correction error is increased by relying only on the projection domain or image domain information.Therefore,this paper designs a dual-domain joint metal artifact reduction network under the framework of generating an adversarial network through the dual-domain prior information constraint.Firstly,the generator is used to deeply mine the data distribution characteristics of polymetallic traces in projection images and the sinusoidality of various voxels in the projection space position,and the back-projection module with back-propagation ability is added to realize the transformation of projection to images and improve the extraction of prior information;then the discriminator is introduced in the projection domain and image domain,respectively,to strengthen the utilization of dual-domain information,further restrict the generator optimization and improve the learning ability of the generator to the detail information;a dual-domain matching simulation data is constructed for network training and performance verification;the experimental results show that compared with the linear interpolation method,the error in the correction results of this method is reduced by more than 50%,effectively removing the metal artifacts in complex metal CT images,especially well eliminates the residual artifacts of metal edges.