Towards The Research Of Low Dose CT Reconstruction Via Sparse Tensor Dictionary Learning

Low dose CT plays an important and indispensable role in such CT scanning cases on which the effective absorption dose is limited clinically,such like the screening of potential lung cancers to the crowd of the smoking.Therefore it is one of the primary hot research fields of advanced CT that how to accurately capture the internal characteristics of the scanned object through using CT while reducing the effective doses simultaneously.Among three key protocols of dose-reduction techniques,lowering the tube-current-time seems a promising candidate and attracts hot blinks of CT researchers' eyes due to its physics similarity to that of regular CT protocol.The main problems of this protocol is that the signal-to-noise ration of raw projection data is rather low and the noise itself follows a mixture model of Poisson and Gaussian distributions,which disables the conventional CT algorithm accurately reconstruct the CT imaging of low dose.Moreover,in such a protocol,how to ultimately collect the X-ray particles penetrating through the scan object is the key research point of detector system design of low dose CT.In this thesis,the preliminary studies of objective validation on the top of the protocol of tube-current-time reduction was performed for the reliable reference to the further optimization of this protocol-based low dose CT.Utilizing primary natures of tubecurrent-time-reduction low dose CT,this thesis conducted the following work:1.Considering the distinct features of X-ray detection and measurement(such as the energy spectrum similarity to that of regular CT and the maximum collection of informational X-ray particles)and grounding the basis of the mechanic driving platform of regular CT,the GOS ultrafast ceramic solid-state scintillator is adopted as the detector array of low dose CT and the parameter-tunable X-ray tube with a rotating anode of tungsten is used to serve the X-ray source;meanwhile,a multi-function anatomical lung phantom is employed as the test object in the low dose CT platform and helps generate the raw project data under a specific case of low dose CT and support the algorithm validation of CT imaging reconstruction;2.Conventional algorithms of CT imaging reconstruction reshape the matrix of projection data into a vector form and inversely compute the vector-form CT images from these vectorized projection data,this makes the structural information hidden in the projection data disappear;therefore in this thesis,a learning approach of sparse tensor dictionary is proposed to utilize the structural information in projection data and further extend itself to the imaging matrix reconstruction of low dose CT;3.Firstly,the traditional digital phantom is used to validate and compare the proposed algorithm of this study with other three state-of-art algorithms and secondly all four algorithms respectively reconstruct the imaging matrixes of the tested lung phantom under the low dose CT scan,the numerical experiments demonstrated that the proposed algorithm of this study is capable of reconstructing the HU imaging of low dose CT.In the study of this thesis,the research highlights are that(1)it is the first time that using the tensor dictionary to denote the mapping relation from the projection matrix to imaging matrix can reserve the structural information hidden in the projection data and imaging,and(2)using the regular CT imaging datasets that have similarities of physics process and anatomical structures to the low dose CT as the training dataset of dictionary learning can minimize the impact of noise to the dictionary learning.