Research On The Itarative Reconstruction Based On Total Generalized Variation For Cone Beam CT

As one of the most advanced imaging technology, computed tomography has been widely used in medical diagnose, industrial non-destructive testing and other fields. Currently in order to reduce radiation dose and increase the efficiency of scanning, reconstruction from incomplete-view projection data has become a hot spot in the research of CT imaging. Total variation regularization method is based on the priori assumption that reconstructed images can be approximated with piecewise constants, so that it can effectively overcome the artifacts and noises in the incomplete-view reconstruction. However, the method could easily lead to details excess smoothing and staircase effect. To solve the problem, researchers have proposed a novel total generalized variation model, which can effectively approximate any order polynomial function, so that it can better maintain the image piecewise consecutive details, and has been achieved preliminary applications in the field of image processing.In this paper, efficient solution strategies have been explored for the cone-beam CT iterative reconstruction based on the TGV regularization model. First, the projection model has been studied for data fidelity term in the iterative reconstruction model, and the parallel algorithm has designed to compute the forward projection and the back projection based on distance driven model. Secondly, for the TGV regularization term, an efficient reconstruction algorithm has been designed based on total generalized variation minimization. Finally, an acceleration strategy based on GPU-cluster platform has been proposed to further improve the computational efficiency of the reconstruction algorithm. The main work is as follows:1. A fast forward projection and back projection algorithm based on three-dimensional distance driven model has been proposed. Distance-driven model has good performance in the CT three-dimensional projection model, but the current computed structure of circulating the projection point is not suitable for parallel computing in the existing computing methods. To solve this problem, an efficient parallel algorithm based on the circulating the detector bins contribution region has been proposed to realize the distance-driven model forward projection and back-projection. The method is based on a three-level parallel structure, which consists of the detector bins, image layer, and the inner layer. By circulating the region of contribution detector bins, it can achieve the model matching and get a good parallel forward projection and backprojection computing architecture. Results using simulation data and real data show that the algorithm can obtain 170 times speedup ratio than the serial algorithm, and maintain higher accuracy than the usual parallel algorithm with the unmatched model.2. An image reconstruction algorithm based on total generalized variation regularization, named TGV-ADM(Total Generalized Variation-Alternating Direction Minimization), is proposed. Based on the theory of sparse image reconstruction and the framework of augmented Lagrange function method, the TGV regularization term has been introduced in the computed tomography and is transformed into three independent variables of the optimization problem by introducing auxiliary variables. Then using the alternating direction method, the TGV regularization term is decomposed into a series of sub-problems that have analytic solutions. In the sub-problems of TGV regularization term, the calculation of the difference matrix is efficiently realized using the FFT technique. In the image f sub-problem, as the calculation for matrix’s pseudo-inverse is too costly, a linear and approximate point technique is applied to make the FFT-based calculation of the analytical solutions in the frequency domain feasible, thereby significantly reducing the complexity of the algorithm. Experimental results with the simulation data and real data both show that in three-dimensional cone-beam CT reconstruction, the reconstruction speed of our proposed algorithm is quite equal to the mainstream TV reconstruction algorithms, and the staircase effect arising by the TV reconstruction can be significantly reduced in the reconstruction results of our proposed algorithm.3. An accelerating strategy using on multiple graphic processing units(multi-GPU) based cluster system is proposed for the iterative reconstruction algorithm. Aiming at the problem that huge computing resource is needed in solving iterative reconstruction algorithm, an accelerating method combining the multi-GPU technique and cluster system technique is designed to accelerate the corresponding method. Based on the data and task characteristics of iterative reconstruction algorithms, strategies for data partition, data communication and internal GPU optimization are proposed to achieve the speedup for the reconstruction algorithm. Experimental results show that the method can improve the speed of iterative reconstruction algorithm, while obtain the same reconstruction quality using a single computer at the same time. The speedup of the reconstruction algorithm upgrades with the increasing number of computing nodes.