Research On X-ray Multispectral CT Imaging And Characterization Algorithms Based On Prior Information

X-ray multispectral CT imaging introduces the attenuation information of objects at multiple energies compared with traditional CT imaging,which can suppress the hardening artifacts of "spectral averaging" and realize the quantitative characterization of material components and microstructures.The technology is significant to the research and performance testing of new materials,new drugs,and new energy.Multispectral CT imaging based on "cocktail party" blind separation avoids the limitations of existing multispectral CT imaging,such as complex models and known spectral prior.However,it still has shortcomings in multispectral artifact removal,noise suppression,and energy orientation of narrow-energy-width spectra,which affects the accuracy of quantitative characterization.Therefore,the thesis focuses on improving the reconstruction quality and quantitative characterization accuracy of multispectral CT based on existing multispectral blind separation methods and conducts research on X-ray multispectral CT imaging and characterization algorithms based on prior information to make multispectral CT easy to implement and apply in quantitative characterization problems.Aiming at the noise amplification problem caused by the ill-formedness of the multispectral blind separation model,an X-ray multispectral CT blind separation algorithm based on Poisson prior is proposed.The algorithm introduces material regularization in the projection domain to suppress noise caused by the ill-conditioned nature of the multispectral blind separation model.Meanwhile,a weighted least squares optimization model with constraint and regularization on the spectral weight vector and thickness vector is developed based on the maximum likelihood principle.The model introduces the Poisson statistical properties of polychromatic measurements.In the framework of block coordinate descent,the nonnegative matrix factorization combined with the Gauss-Newton algorithm is used to perform alternate iterations to solve the model effectively according to the differences in the number scale,location,and constraints of variables to be solved.In addition,an efficient initialization strategy is given for the model's nonconvexity.Simulation and practical experiments indicate that the algorithm can effectively suppress the noise of decomposition projections and improve the quality of narrowenergy-width reconstructed images and the accuracy of attenuation coefficients.Aiming at the problem that projection edges are over-suppressed due to the non-strict smoothness of material projection data,a blind separation algorithm of X-ray multispectral CT based on image sparse prior is proposed.The algorithm introduces the imaging system matrix into the X-ray multispectral blind separation model to achieve direct inversion of material images.This can not only avoid the introduction of indirect errors but also convert the projection-domain regularization into the image-domain regularization with edge-preserving properties so that the edge and detail information of the image can be preserved while reducing noise.Based on the maximum likelihood principle,a constrained regularized minimization model for the spectral weight vector and volume fraction vector is established,and an alternating optimization algorithm is used to solve it.The optimization transfer principle is adopted to solve the model,which transforms large-scale nonlinear optimization into a series of independent small-scale quadratic programming.Then the algorithm efficiency is improved by parallel computing.Experiments indicate that the narrow-energy-width images decomposed by the algorithm not only have fewer hardening artifacts but also have more accurate attenuation coefficients.The CT image sequence obtained by multispectral separation reconstruction enables quantitative characterization of material components.However,relevant parameters need to be adjusted repeatedly to ensure the characterization accuracy,resulting in poor practicability.Therefore,given the adaptability of prior parameters,a penalized weighted least squares model for component characterization is constructed under the physical mechanism of multispectral CT component characterization.Then,drawing on the idea of the LEARN network in the image reconstruction field,the gradient descent formula of the model is expanded into an iterative deep network,and the prior parameters are automatically adjusted through the learning of network parameters.The network builds on traditional iterative methods and can consider the consistency of data before and after characterization.Only a small dataset is required for network training.Simulation and practical experiments indicate that the method has obvious advantages in image noise suppression,and can also accurately characterize materials with similar attenuation characteristics.Meanwhile,the method has a certain fault tolerance ability.It can effectively characterize the narrow-energy-width images obtained by several multispectral separation reconstruction algorithms.The proposed method reduces the difficulty of multispectral CT image characterization and improves the accuracy of quantitative characterization.