Based On The Cuda Gpu Accelerated Iterative Reconstruction Algorithms

The use of x-ray computed tomography (CT) scanners has become widespread in both clinical and industrial areas. As one of the core technologies of the computed tomography, CT reconstruction algorithm is always the research focus. It is mainly classified as analytical reconstruction algorithm and iterative reconstruction algorithm. Analytical reconstruction algorithm currently has been widely used because of fast implementation and image quality. But it needs to measure a large number of projections distributed uniformly over180°or360. However, this condition is not always satisfied because of practical constraints due to dose limited, or data missed when x-ray passes through high density region. In these cases, the analytical reconstruction algorithm will lead to conspicuous artifacts. On the contrary, an iterative reconstruction algorithm is usually employed to get better reconstructed images.Iterative reconstruction algorithm is applicable to all forms of projection data, even though projection data is not sampled enough. However, Iterative reconstruction algorithm is not commonly used, due to its computational burden (time consuming). So facilitating the implementation of iterative reconstruction algorithm, as a hot issue of study, has great contribution to the field of theory and practical application.In this paper, we proposed an acceleration method used for Iterative reconstruction algorithm based on GPU with high-performance parallel computing ability. The major research contents are as follows:1) We accelerate Simultaneous Algebraic Reconstruction technique (SART) on the platform of CUDA(Computer Unified Device Architecture). Two core techniques are proposed to fit SART into the CUDA architecture:(1) a ray-driven projection along with hardware trilinear interpolation, and (2) a voxel-driven back-projection that can avoid redundant computation by combining CUDA shared memory. We utilize the independence of each ray and voxel on both techniques to design CUDA kernel to represent a ray in the projection and a voxel in the back-projection respectively. Thus, significant parallelization and performance boost can be achieved.2) We study TV reconstruction which is based on the minimization of the image total variation. Combing the gradient of TV and SART algorithm, we propose a new algorithm named TV-SART algorithm and accelerate it on the platform of CUDA. Simulation experiments and practical tests show that the TV-SART suppresses noise well and smoothes reconstructed image. 3) On the basis of a single-GPU-accelerated reconstruction, we propose a Multi-GPU parallel acceleration method.This method takes full advantage of multiple GPU core computing power to further improve the reconstruction speed. In addition, it avoid frequent data copying between host and device, as single GPU is lack of enough memory in case of a large amount of data.. Therefore, the data transmission time is reduced.