Data-Model Coupling Driven Low-Dose CT Precision Imaging

X-ray computed tomography(CT)has great advantages in obtaining the internal anatomical structure of patients,making CT imaging widely used in clinical diagnosis and treatment.However,excessive X-ray radiation increases the risk of cancer for patients,especially in CT perfiusion(CTP)imaging applications.Currently,the simplest solution to reduce radiation dose is to directly decrease the tube milliamps or the scan angles.However,the reduction of radiation dose will inevitably destroy the consistency of the original sinogram data and further affect the quality and speed of CT image reconstruction.How to reduce the X-ray radiation dose while ensuring fast and high-quality CT image reconstruction has become the most urgent problem in clinic.At present,widely used low-dose CT(LdCT)imaging methods can be broadly divided into two categories:model-driven methods and data-driven ones.The model driven methods can be further divided into two subclasses:Analytical filtered back-projection(FBP)and Model based iterative reconstruction(MBIR).The advantage of FBP is that CT image reconstruction can be performed fast,but the corresponding image reconstruction quality cannot be guaranteed.The MBIR can achieve better image quality than the FBP,but properly designing prior knowledge and adjusting hyperparameters are not simple tasks.More importantly,since the MBIR optimization process involves multiple forward and backward projection operations,the reconstruction speed will be much slower than that of FBP.In recent years,data-driven LdCT imaging methods have been extensively studied with the development of deep learning(DL)technique.Promising processing results have been achieved for LdCT imaging tasks.However,at present,most data-driven LdCT imaging algorithms directly use standard DL modules and perform denoising in the sinogram and/or image domains,ignoring the intrinsic mechanism of CT imaging;this greatly limits the potential advantages of data-driven algorithm for the LdCT imaging problems.In this thesis,focusing on the difficulties of existing LdCT imaging methods,the author has conducted in-depth researches on data-model coupling driven algorithms for fast and promising LdCT imaging.To sum up,the main works of this thesis are as follows:(1)Combining the traditional MBIR algorithm and deep learning technique,a data-iterative model coupling driven parameterized plug-and-play ADMM(Alternating Direction Method of Multipliers)framework is proposed.Compared to traditional methods,the new framework uses some leanable network parameters to replace the hyperparameters in the ADMM iterative format.In addition,a standard neural network is used instead of the design of the regular term;then the modified ADMM solution format is unfolded into a flattened network,which can be optimized using backpropagation and stochastic gradient descent methods.Finally,the optimized model is used for CT image reconstruction.The new algorithm addresses the problem of prior knowledge design and hyperparameter selection in iterative CT reconstruction algorithms.(2)the data-iterative-model coupling driven parameterized plug-and-play ADMM strategy is extended from the 2D LdCT image reconstruction task to the 3D LdCT perfusion deconvolution application,thus solving the problems of prior knowledge determination and hyperparameters optimization in the calculation process of perfusion residual function,which further verifies the superiority of this strategy in LdCT imaging.(3)Combining the analytical FBP algorithm and DL technique,a Radon inversion unified framework based on data-analytical model coupling driving is proposed.The algorithm draws on the reconstruction paradigm of analytical FBP algorithm,and constructs a filter operator’along the angular direction in the sinogram domain and a sinusoidal back-projection operator by using DL technique.Among them,the filter operator is constructed using the fully connected technique;and the sinusoidal back-projection operator is constructed according to the relationship between the reconstructed image and the sinogram.Finally,to enhance the overall flexibility of the network,a post-processing sub-network is added after the sinusoidal back-projection operator.The algorithm is overally optimized by training a large amount of data,and achieves fast and robust LdCT accurate image reconstruction.The experimental results show that the proposed data-model coupling driven LdCT imaging methods have superior performance in both qualitative and quantitative measurements.