Research On Iterative Reconstruction Algorithm For Cone-beam CT And Its GPU Cluster Acceleration

Computed tomography(CT) is an advanced non-contact and non-destructive testingtechnology, which has now been widely applied to medical diagnosis, industrial non-destructivetesting and security inspection, and other fields. In recent years, regarding the application needsof the large objects with special structure, low-dose and low-cost scanning, as well ashigh-efficiency detection, due to its unique advantage of solving the incomplete projection datareconstruction, iterative reconstruction algorithm becomes a research hotspot in the field of CTreconstruction.Focusing on the characterization of projection matrix, iterative reconstruction algorithm andits acceleration technology, this paper develops study and achieves following research results:1. It proposes projection matrix characterization method based on the finite element modeland Radon operator. It estimates the heterogeneity of substances within the voxel through finiteelement model, estimates the intensity of ray and voxel through Radon operator, and optimizesthe new model by means of discrete linear difference. As this method comprehensively reflectsthe influencing process of the ray and the voxel, the characterization of the projection matrixbecomes more precise. The experimental results show that the proposed model effectivelyimproves the reconstruction quality of the iterative reconstruction algorithm.2. Based on highly precise characterization of projection matrix, it proposes an adaptiveregularization iterative reconstruction algorithm on the basis of a sparse constraint: the AR-SART-CG(Adaptive Regularization-Simultaneous Algebraic Reconstruction Techniqueis-Conjugate Gradient, AR-SART-CG) algorithm. The algorithm adopts a kind of Lagendijk type ofRegularization Strategy to construct optimization, respectively uses local variance, noiseestimation, and image energy estimation to adaptively adjust the parameters weighted diagonalmatrix and global regularization, and respectively applies SART algorithm and conjugategradient method to solve optimizations of fidelity term and constraint term. Since that thealgorithm can adaptively adjust the weight of constraints, it is possessed of strong robustness.The experimental results show that the AR-SART-CG algorithm can better balance and preservethe relations between picture edge and smooth noise.3. It designs and realizes a GPU cluster generic acceleration platform: GMatrixCloud,researches and develops platform of message passing mechanism based on Matlab developmentenvironment and CUDA kernel function interface, and enables the platform to effectivelyintegrating the advantages of cluster and GPU acceleration. Besides, it also puts forward the twoparallel speedup strategy suitable for GPU clusters-the AR-SART-CG algorithm, and thisstrategy reconstructs image according to slice division, forms a mixed-granularity two parallelcomputing model with multi-slice coarse-grained parallel among the cluster nodes and singleslice of fine-grained parallelism within GPUs, adopts orthographic parallel technique amongnodes based on binary tree reduction and parallel technology within nodes based on sliceprojection index table to further optimize the two-parallel strategy, applies time feedback cost function to estimate the computing power of nodes and accordingly allocate computing tasks,and then achieve effective load balancing. The experimental results show that when two-parallelstrategy based GPU cluster obtains the same quality as the serial algorithm, it greatly improvesthe reconstruction efficiency.