Research On Key Techniques For Spectral CT Reconstruction Based On Sparse Regularization And Deep Learning

Computed tomography(CT)has been widely used in medical diagnosis,industrial testing,and security inspection since it was invented in the 1970 s.CT can noninvasively generate the attenuation coefficient of the scanned object via the image reconstruction algorithm using the projections collected from different scanning angles.Different from the traditional CT that uses single X-ray spectrum,spectral CT applies two or more X-ray spectra to scan the object and can obtain more various material information by exploring the difference of attenuation coefficients of the scanned object under different energies.Spectral CT effectively improves the ability of X-ray imaging in material identification and analysis,and becomes the development trend for the X-ray imaging.Spectral CT imaging can be achieved by designing specific scanning strategy or using the detector with energy resolving capability.Among them,dual-energy CT and photon couting detector-based CT are the main realization methods for spectral CT,and have attracted much attention and developed rapidly in the field of X imaging.However,the inverse problem of spectral CT reconstruction has significant illness,leading to severe noise boosting in material decomposition.The restrictions of scanning circumstance and photon counting technology further increase the singularity of spectral CT reconstruction,limiting its application scope and technical advantages.Designing efficient reconstruction algorithms for the ill-posed problems in practice is an active research topic in spectral CT imaging and has great significance in theoretical study and practical applications.This thesis focuses on the study of material decomposition and image reconstruction in spectral CT,and does researches on the bottleneck problems of different implementations: noise boosting of material decomposition,the incompleteness of projection sampling,and the low efficiency of photon counting.The main contributions are as follows:1.A direct material decomposition method is proposed based on total variation(TV)and block matching 3-dimension(BM3D)frame.The nonlinear solution of polychromatic projection model and the estimation errors of X-ray spectrum degrade the ability of traditional algorithms to decouple the material information in the projections of different energies,leading to the increase of image noises and the decline of decomposition quality.To solve this problem,the block matching frame and total variation regularization model is applied to explore the nonlocal self-similarity and local sparsity of basis material images,respectively.At the basis of polychromatic projection model,a direct material decomposition model is established based on the joint TV and BM3 D regularization model.Alternative direction method is applied to solve this model.Digital and real data experiments validate the capability of the proposed method in noise suppression and structure preservation during material decomposition.Compared with the single TV and BM3 D regularization methods,the proposed method reduces the noise standard deviation(SD)of basis materials to below 20% and 40% of the original.2.A Butterfly network is designed to perform image-domain material decomposition.The noises of CT images under different energies are accumulated and magnified via the coefficients of decomposition matrix,reducing the quality of basis materials in spectral CT.This paper analyzes the image-domain decomposition model and aim to solve the ill-posed problem of material decomposition by utilizing the nonlinear mapping capability of neural network.Based on the relationship between dual-energy CT images and basis materials,this paper designs a double-entry double-out Butterfly network.The crossover architecture of the designed network implements the information exchange between the two material generation pathways.The entire network is divided into three stages: characteristic extraction,information exchange,and characteristic formation.Digital and real data experiments validate that the proposed method is capable of suppressing the boosting noises in material decomposition.Compared with the matrix inversion and iteration methods,the proposed method reduces the noise SDs to below 10% and 20% of the original.3.An incomplete angle spectral CT reconstruction method is proposed by using the prior knowledge of complementary support set(Pri-CSS).Most spectral CT systems require the full-scan projection data,restricting its applications to the imaging areas with incomplete scanning angles.To solve this limitation,this paper designs a scanning strategy for dual-energy CT with one half-scan plus a second limited-angle scan.To solve the incomplete angle problem under this scheme,this paper analyzes the key information of the necessary and sufficient condition for the accurate reconstruion of TV minimization model.Inspired by the consistency of complementary support set between high-and low-energy CT images,this paper extracts the Pri-CSS from the gradient image of the first half-scan CT image to promote the second limited-angle CT reconstruction.Pri-CSS are incorporated into TV regularization model in the form of constrains.Alternative direction method is applied to iteratively solve the modified optimization model.Digital and real data experiments show that the utilization of Pri-CSS largely reduces the limited-angle artifacts and the proposed method reconstructs dual-energy images with high quality.Compared with traditional TV-based method,the proposed method reduces the reconstruction errors under different scanning angles by about 30% on average for real data.4.A novel imaging configuration with dual quarter scans scheme and its reconstruction method are proposed for spectral CT.To further reduce the scanning angle and radiation dose and explore the possibility of dual-energy CT in half-scan,this paper designs a dual quarter scans configuration,which consists of two 90-degree scanning arcs with different energies.To solve the dual limited-angle problems,this paper studies the characteristics of image artifacts with positive and negative values under dual quarter scans,and found that the limited-angle artifacts of dual-energy CT images are complementarily distributed in image domain.Inspired by this finding,a fusion CT image is generated by integrating the limited-angle dual-energy CT images under dual quarter scans,which enhances the real image information and suppresses the false aritfacts.Furthermore,utilizing the capability of neural network in the modeling of complicated mapping relationship,a novel Anchor network with single-entry double-out architecture is designed in this work to yield the desired DECT images from the generated fusion CT image.Digital and real data experiments show that the proposed method obtains the almost same reconstruction and decomposition quality with the full-scanning results.Compared with the iterative and network methods,the proposed method reduces the reconstruction errors of highand low-energy CT images to 30% of the original and increases the signal to noise ratio by about 10 d B.5.A multi-energy reconstruction method is proposed using tensor nonlocal similarity and spatial sparsity regularizations for photon-counting-based spectral CT.Due to the low counting rate and spectrum distortion of photon-counting detector,the projections under different energies have great singularities,which degrades the reconstruction quality of multi-energy CT.To solve this problem,this paper analyzes the prior characteristic among interchannel images,and a three-order tensor is formulated by matching the nonlocal similar patches from the spectral-spatial domains.Intrinsic tensor sparsity regularization model is applied to explore the low-rank and sparsity of the generated tensor.TV regularization model is also introduced to depict the sparsity of channel images.A multi-energy CT reconstruction model is established by incorporating the two abovementioned regularization terms.Alternative direction method is utilized to solve the established model based on a flexible framework.Digital and real data experiments indicate that the proposed method can reduce the influences of detector physical restrictions and shows obvious advantages in noise suppression and detail preservation.Compared with the tensor low-rank-based method,the proposed method reduces the reconstruction error of digital data by about 30% and the noise SD of real data by about 50%.