Research On The Imaging Technologies For Short-scan In Cone-beam CT

Cone-beam computed tomography (CBCT) has been widely used in industrial nondestructive testing and medical inspection due to the characteristics of high imaging speed, high spatial resolution and non-destructibility. Compared to the 360° scan in cone-beam CT, the short scan can achieve a fast and low dose scan by irradiating the object within π plus cone angle. However, the geometric calibration for the short scan system is quite sensitive to error as the lack of projection data, and the reconstructed image is affected greatly by the uniformly redundant projection in the short scan.In view of the problems mentioned above in the short scan,this paper mainly focuses on the geometric calibration,reconstruction algorithm and its acceleration,respectively. The major achievements are as follows:(1)The lack of projections in short scan makes the geometric calibration quite sensitive to the inputs, which is the projection location of the calibration object’s center, and we call it the center projection. The calibration object is invariably a ball.There is a big error when the center projection is approximately estimated by the center of the support of ball’s CB projection, or the centroid of the intensity values inside the support, in the existing methods.Aiming at this problem, a theoretically exact method is proposed to calculate the location of the center projection.Firstly, two models are established for (i) the relation between the geometric parameters and the shape parameters of ball’s CB projection and (ii) the relation between the geometric parameters and the center projection.Uniting the equations, the center projection can be described only by the shape parameters of the ball’s CB projection, which are obtained by the least squares fitting to the boundary point of the projection.The experimental results show that our method has obvious advantage in precision with a good robustness, compared with the existing methods.(2) The reconstruction image is disturbed by the uniformly redundant projection in short scan.Aiming at this problem, a selective backprojection filtration(S-BPF) is proposed. Based on the analysis of the redundancy of the projection, a selection principle is derived by considering the direction of PI line and the relative position between the source and PI line. With this principle, the scope of the image is determined in advance and the reconstruction image is free from the redundant data with a less calculation of the backprojection.The experimental results show that the algorithm can avoid the influence of the redundant data effectively and achieve a high quality reconstruction.Compared with the existing algorithms using data rearrangement or data weighting in projection to deal with the redundant data, the proposed algorithm has a less cost in time and memory.(3)The GPU acceleration of reconstruction algorithm has shown a good performance, and the communication delay is the bottleneck of the acceleration. Aiming at this problem, a strategy is proposed to overlap the time in communication. Based on the analysis for the parallel performance of S-BPF algorithm and the characteristic of the CPU-GPU platform, the parallel execution is realized between (ⅰ)the CPU calculation and the data transform in the hard disk-memory by multi-thread technique, (ⅱ)the CPU calculation and the data transform in memory-device memory by the CUDA asynchronous technique, (ⅲ)the GPU calculation and the data transform in memory-device memory, device memory-device memory by the multi-stream technology and asynchronous technology. The proposed strategy hides the data transform among hard disk, memory and device memory by overlapping it with the calculation in GPU to improve the efficiency of the reconstruction. The result shows that the proposed strategy can save a half of the time cost, based on the existing strategies.