Research On Noise And Artifact Suppression Algorithm Of Low-dose CT Image

Computed Tomography(CT)technology has an irreplaceable role in the current medical diagnosis field.During the three-year period of the novel coronavirus,CT scans became the preferred solution for pneumonia screening,due to their advantages of clear tomographic images and high resolution for human lung detection,and also made more people aware of CT technology.With the development of CT equipment and CT reconstruction algorithms,the application field of CT scanning in medical diagnosis has been expanding and the frequency of CT use continues to increase.However,due to the problem of ionizing radiation from X-rays,research scholars and CT manufacturers have been focusing on Low-dose CT(LDCT)imaging technology.Low-dose CT scanning technology is favored by patients and physicians,because it can reduce radiation damage to patients.Currently,low-dose CT scans are applied to clinical screening and diagnosis gradually,however,low-dose CT images contain a large amount of mottle noise and streak artifacts,which degrade the image quality and affect radiologist’diagnosis of the disease.In this paper,three image domain post-processing algorithms based on partial differential equation(PDE)diffusion model approach and the data-driven deep learningbased approach are proposed,which focus on suppressing the noise and artifacts of low-dose CT images.The main works are as follows:(1)A novel anisotropic diffusion fourth-order PDE algorithm is proposed for low-dose CT image denoising.Unlike the fourth-order PDE anisotropic diffusion model which uses gradient magnitude to determine the diffusion coefficient,in the novel diffusion model,the residual local energy with a weight factor combined with the image gradient magnitude is used to jointly determine the diffusion coefficient.The residual local energy as a texture and detail detection operators,overcomes the deficiency of using gradient magnitude to detect image edges and ignoring the protection of texture and details.In addition,in order to make the processing results closer to normal dose CT(NDCT)images,avoid excessive smoothing and reduced block effects,a fidelity term is introduced in the diffusion model.The experimental results show that the proposed algorithm protects the texture details and suppresses the block effects of low-dose CT images compared with the anisotropic diffusion fourth-order PDE model.Furthermore,the proposed model effectively suppresses the mottle noise and streak artifacts to improve the quality of low-dose CT images compared with other excellent algorithms.(2)To solve the problem of blurring in edge and texture easily in the denoising process of LDCT images,a fourth-order PDE diffusion model with adaptive Laplace kernel is proposed.The guided filter with edge protection effect is added as the fidelity term in the objective function;The gradient magnitude and grayscale variance are used in the diffusion function to form the edge and detail detection operators to protect the edge and detail information of LDCT images;In the discrete implementation of the diffusion model,the adaptive Laplace operator with stronger edge protection capability is used instead of the traditional Laplace operator.Experiments show that the proposed fourth-order PDE diffusion model with adaptive Laplace kernel achieves excellent edge protection and noise suppression in low-dose CT images.In terms of visual effects and objective evaluation indexes,the proposed model performs better compared with other excellent algorithms for processing low-dose CT images.(3)A deep convolutional dictionary learning algorithm with no noise parameter is proposed.The convolutional dictionary learning method has a clear physical meaning,the advantage of shift-invariant property.The deep convolutional dictionary learning algorithm(DCDic L),which combines deep learning and convolutional dictionary learning,has excellent denoising performance on Gaussian noise.However,using DCDic L for low-dose CT images does not get satisfactory results.The DCDic L algorithm needs to input noise intensity parameter in the network module,while in practical situations,clinical low-dose CT images do not have noise statistical property and noise intensity parameter is difficult to estimate.Therefore,we modify the input network module so that only input low-dose CT images and no need noise parameter,which enhances the convenience and applicability of the algorithm.Meanwhile,in the subnetwork module for solving the prior information of the convolutional dictionary,the dense convolutional network Dense Net121 is used instead of the shallow convolutional network,which can extract the prior information of the convolutional dictionary more accurately.Then,obtaining a more accurate convolutional dictionary and improving the denoising performance of the algorithm.The DCDic L algorithm uses L1 loss,which tends to make the processing result excessive smoothing.We add MSSIM loss to enhance the detail retention ability of the algorithm.The experimental results show that the proposed algorithm enhances the denoising performance for low-dose CT images with reduced input parameter,and effectively improves the quality of low-dose CT images.