Low-dose CT Imaging Via Deep Learning And High-dimensional Tensor Modeling

Computed Tomography(CT)takes advantage of the penetrability of X-rays to scan the human body in a non-invasive and non-open injury manner,and then relies on detectors to collect projection data.The projection data are used to reconstruct the image of internal anatomical structure.However,when X-rays pass through the human body,the interaction between X-photons and atoms in the body will cause ionizing radiation damage to the human body.Therefore,it is the long-term goal of researchers in related fields to reduce the radiation dose of CT to the maximum extent under the premise of obtaining the image quality that meets clinical needs.In clinical practice,the use of Low-mAs scan mode is an effective way to reduce CT radiation dose.However,in the low-mAs scanning,the collected projection is contaminated by measurement noise,which results in serious noise and artifacts in the reconstructed image.Many advanced algorithms have been proposed to address the problems of low-dose CT reconstruction.It is worth paying attention to the low-dose CT challenge jointly held by Mayo Clinic and other institutions in 2016.In this challenge,the low-dose CT reconstruction methods based on deep learning technology showed a trend of being superior to the traditional iterative reconstruction algorithm in terms of reconstruction effect and computational efficiency,marking that the research of low-dose CT imaging has entered a new stage of intelligent imaging.On the other hand,sparse-view CT(SCT)is one of the most promising protocols for ultra-low-dose scanning.However,as the sampling of projection data decreases,the SCT reconstruction will no longer satisfy Shannon's sampling theorem,and the image is contaminated by serious artifacts.While many SCT algorithms have been proposed,they were usually designed for circular imaging geometry.Thus,the lack of research on the sparse-view Helical CT(SHCT).In response to the above-mentioned key issues of ultra-low-dose CT imaging and SHCT imaging,this thesis proposes the following works:(1)An iterative residual-artifact learning convolutional neural network(IRLNet)is proposed to solve the problem of ultra-low-dose CT image reconstruction with the low-mAs scanning.The IRLNet model adopts the residual learning technology and wavelet method to separate complex artifacts from normal anatomical structures,and then removes artifacts and noise in the image through image iteration process.The results show that IRLNet can effectively suppress the noise and artifacts of ultra-low-dose CT images.Besides,this study also proposes the IRLNet-based iterative reconstruction mode to further improve the performance of ultra-low-dose CT reconstruction.(2)A sorting view-by-view backprojection tensor learning network(SVBT-Net)is proposed to address the limitations of the IRLNet model.S VBT-Net can avoid information compressing caused by the summation step in the filtered backprojection(FBP)algorithm,and therefore allows the SVBT-Net adaptively processing lossless information of the data.The results reveal that the SVBT-Net learning framework obtains obvious improvement over the image domain,the projection domain,and the hybrid projection-image domain learning frameworks.(3)To address the problem of ultra-low-dose SHCT reconstruction,the three-dimensional(3D)anatomical structure sparsity of the SHCT image is analyzed in this study.Based on the analyses,we propose a tensor decomposition and anisotropic total variation regularization model(TDATV)for SHCT iterative reconstruction.The results reveal that SHCT could serve as a potential solution for reducing HCT radiation dose to ultra-low level by using the proposed TDATV model.