Developing New Methods For Low-dose CT Imaging With Backprojection Data Analyses And Temporal Average Prior

Computed tomography(CT)is a nondestructive imaging modality in vivo based on X-ray imaging technique,which has many advantages including,fast scanning speed,high spatial resolution,and the ability of accurately delineating the anatomical information of patients.Due to the mechanism of X-ray imaging,radiation is an inevitable problem in CT imaging.Reconstructing high quality images meeting clinical requirement from data acquired with low-dose scanning protocol has long been the persuit of physicians,researchers and manufacturers.As the standard reconstruction algorithm used in commercial CT machines,the filtered backprojection(FBP)algorithm has gained wide applications in industry.However,FBP is susceptible to data corruptions.When the radiation dose is reduced,the images reconstructed by FBP will contain severe noise and artifacts,which reduce the diagnostic accuracy.To deal with this problem,previous studies often adopt strategies including projection preprocessing,image postprocessing and iterative reconstruction,which are based on property analysis of projections or images.However,to date,no matter targeting the analytical or iterative reconstruction algorithm,there has been rare studies analyzing and exploiting the intermediate data during CT reconstruction.This paper aims at studying the most important two types of algorithms--analytical and iteration algorithms for low-dose CT imaging.For the analytical type algorithm,the study analyzes properties of the BP-Tensor(Backprojection tensor)during FBP reconstruction,specifically designs an algorithm to process the BP-Tensor based on its properties,exploits its applications in low-dose CT imaging and discusses its potentials in other CT imaging tasks.For the iterative type algorithm,the paper focuses on low-dose dynamic CT imaging problem such as CT perfusion or four dimensional CT,exploits the naive correlation between the intermediate reconstructed image and the prior average image,and proposes a general iterative reconstruction model for this type of imaging modality.Specifically,the works of this study include the following aspects.(1)Analyzing the BP-Tensor during FBP reconstruction.The backprojections of each view of the filtered sinogram were stacked into a three order tensor,and then sorted in its third dimension by reordering its mode-3 vectors.It was discovered that slices of the sorted BP-Tensor contain structures similar to that of the image,and structures in different slices are correlated.BP-Tensors of patient data from different body parts also express this phenomenon.Properties of the BP-Tensor,including structural self-similarity,tensor sparsity and noise statistics were analyzed,and a new reconstruction framework based on BP-Tensor processing was proposed.The proposed framework not only considers the statistics of the data,but also incorporates structures of the image,and therefore is superior to conventional frameworks.Lastly,reconstructions of low-dose CT data by the proposed framework and other conventional frameworks,including projection preprocessing,image postprocessing and iterative reconstruction,with the same regularization method were compared.It was demonstrated that the proposed framework could obtain results better than other frameworks.(2)Specifically designing an algorithm named TP-tSVD(Tensor processing-tensor singular value decomposition)to denoise the low-dose BP-Tensor based on its properties.The TP-tSVD algorithm uses the penalized weighted least square(PWLS)model to incorporate noise statistics into the objective function,the tensor singular value decomposition(tSVD)as the regularization term to exploit tensor sparsity,and a full-band-patch block matching strategy in each iteration to utilize structural self-similarity.Experiments were conducted using real scanned physical phantom data and patient data with different dose levels.The results suggest that the proposed TP-tSVD algorithm is superior to all competing methods under other reconstruction frameworks.(3)For low-dose dynamic CT reconstruction,analyzing the correlation between the prior average image and the average of image sequence in current iteration,and proposing a general iterative reconstruction model named temporal average prior(TAP)model based on the analysis.The traditional average image assisted reconstruction model often enforce each frame in each single iteration to approach the prior average image,which might introduce mixed temporal information into the reconstructed image because of its strict hypothesis.The proposed TAP model enforces the average of the current image sequence to approach the prior average image,and therefore avoids the introduction of mixed temporal information.Furthermore,the TAP model was formed as a general model based on the plug-and-play theory,where arbitrary denoisers could be used as the regularization term.Therefore,existing algorithms for dynamic CT reconstruction could be improved by integrating them into the TAP model.The low-dose CT perfusion data and the 4D-CBCT data were reconstructed in the experiment.The results demonstrate that,using the same regularization term,the TAP model acquires noticeable gains over the single-frame reconstruction method and the average image assisted reconstruction method.In a word,this paper deeply investigated the CT reconstruction procedures with a new viewpoint.It was demonstrated that by analyzing properties of data during reconstruction and designing algorithms based on the analysis,optimal results could be obtained.The analysis of BP-Tensor,and the proposition of the BP-Tensor processing framework and the TP-tSVD algorithm revisited the classical FBP algorithm with a new insight,providing a heuristic perspective for algorithm development.The TAP model provides a simple,efficient and general model for dynamic CT reconstruction,which might improve many existing single-frame based dynamic CT reconstruction algorithms.It is worth noting that the study about the BP-Tensor provides CT imaging a new processing target within new data domain.The BP-Tensor processing based reconstruction is a new framework for CT image reconstruction.