Research On Artifact Correction For Industrial Cone-beam CT And Its Multi-GPU Cluster Acceleration

Industrial cone-beam computed tomography(CT) is an isotropic radial detecting technology which is widely used in nondestructive testing. However, the differences between the real and ideal CT system result in the artifacts in the reconstruction images which degrade the image quality of the CT system. The improvement of accuracy and speed for artifacts calibration is one of the hotspots and difficulties in CT field. Therefore, researches on rapid and efficient artifact reduction methods have important significance in the development of CT technology.Aiming at the problems of artifact reduction methods in cone beam CT, this paper mainly study on geometric calibration, metal artifact reduction(MAR) and the accerleration of those methods. The major achievements cover the following aspects:(1) Aiming at the problem that part of the geometric parameters have relatively low precision, a self geometric calibration method based on the symmetry of projections is proposed. In this method, five geometric parameters that determine image quality, voxel size and image location are chosen for calibration according to the relationship of cone-beam CT. The in-plane rotation angle of the detector is firstly determined by the structure similarity index of two reconstruction images through the interval subdividing based method. Then the rest four parameters are divided into two groups, and a cost function is established by calculating the mean square errors of symmetry of projections in the mid-plane, the left four geometric parameters is obtained in two steps by minimizing the cost function during which the golden section search method and NM-simplex method are used according to the theory of alternating direction method. The results of experiment show that the presented method makes an outstanding performance in reducing the geometric artifacts and improves the precision of two geometric parameters that have relatively low precision in the compared method and the spacial resolution of reconstruction image.(2) Aiming at the problem that the interpolation based method performs poor in reducing the artifact caused by multiple metallic objects, an exponential formed correction model is proposed for MAR method. In this method, metal regions are directly segmented from projection domain by a simple threshold, then a exponential formed correction model is established for projections in metal regions. The model contains three parameters that are threshold, calibration magnitude and calibration order, and correction is done by adjusting parameters of the model. The optimal solution of the parameters is achieved by NM-simplex method that makes the gray entropy of the reconstruction image minimum. The results of experiment show that the presented method significantly reduce the radial and streak artifacts due to multiple metallic objects, meanwhile it restrains the beam harden phenomenon, and improves the signal-to-noise ratio of reconstruction image in regions of interest.(3) Aiming at the problem that the image reconstruction process during the artifact reduction becomes time-consuming when faced with the large data sets, a three-level parallelism method for reconstruction algorithm using multiple graphic processing units(multi-GPU) based cluster system is proposed. In this method, the cluster system is divided into three levels according to its stucture, that is layer between nodes, layer between GPUs and layer of threads in single GPU. Then the recostruction tasks are assigned to three levels according to its characteristics and corresponding parallel strategies are designed for each. Compared with the state-of-the-art GPU accelarating technology, multi-GPU cluster system combines the multi-GPU technology and cluster system technology and sufficiently takes the advantages of both. The results of experiment show that the presented method obtains the same image quality as the single computer,meanwhile it increases the speed of artifact calibration under a large scale of data sets, and with the increasing number of computing nodes or GPUs per nodes, the speedup improvement of artifacts calibration is notable.