Fast Iterative Reconstruction Algorithm Of Cone-beam CT From Sparse-view Data Based On CUDA

Three-dimension cone-beam CT has become a popular research direction in the field of computer tomography.Compared with two-dimension CT imaging,cone-beam CT,widely used in clinical applications,has many advantages,such as short scan time,low partial volume effect,high vertical resolution and high radiation efficiency.In view of the impact of X-ray radiation on human cancer,leukemia and other genetic diseases,reducing the dose of radiation received by the patients become one of the main objectives of CT examination.There are two ways to reduce the radiation dose,one is reducing the tube voltage or tube current,the other is using sparse-view scan.Different from the former,sparse sampling can gain high-quality projection data,decrease the motion artifacts caused by the patient’s involuntary movement and decrease total radiation dose by shorten scan time.But the reconstruction from sparse-view data is a serious ill-posed problem.In this situation,the conventional analytical methods based on projection geometry can introduce many strip artifacts in imaging results.And by the general iterative algorithms such as ART or EM method,we also cannot gain satisfactory reconstruction images.By adding prior information to the objective function of the EM algorithm,the model of Maximum A Posteriori(MAP)can effectively suppress the artifacts caused by sparse sampling and obtain better reconstruction results.Being an iterative method,the EM algorithm of three-dimension cone-beam CT has high time complexity so that the CPU does not have enough capacity to support the iterative optimization process of EM algorithm.The operations of projection and back-projection are the main bottlenecks that limit the speed of the EM method.In recent years,the optimization strategy of objective function using GPU,with strong parallel computing power,has become an effective solution.But not all program transplanted to the GPU platform can get satisfactory speedup ratio.The acceleration effect involves the choice of algorithmic model,GPU hardware resources,auxiliary space selection and engineering programming skills and so on.In this thesis,the projection model and back-projection model suitable for parallel are selected.And through experiments,we discuss the acceleration effects of projection and back-projection speed by different acceleration measures,include fixing sampling number of projection,using sum reduction based on shared memory,using cylinder box,using texture memory and so on.Then we optimize the implementation of EM algorithm.MAP algorithm constrained by non-local prior has been applied to two-dimension CT reconstruction.It can suppress the strip artifacts caused by sparse sampling and generates high-quality reconstruction image.However,the applications of non-local method in three-dimension cone-beam CT are limited by its massive calculation.In order to solve this problem,we propose an algorithm named ordered sub-window search.The large search window is divided into several subsets.In each iteration,one of the subsets is chosen for the calculation of prior function.Experiments show that the proposed method ensures the quality of the reconstruction image similar to the results with traditional non-local algorithm,the computation speed for each iteration is largely increased.