基于Katsevich算法的CBCT多切片重建的优化设计与GPU实现
发布时间：2018-08-22 19:24
【摘要】：螺旋CBCT当前广泛用于临床检测、射线探伤等行业,随着电子信息、传感器的快速发展及应用需求的提高,对其重建质量、成像速度及资源消耗都提出了更高的要求。作为精确重建的Katsevich算法能够保证成像质量,但因计算复杂、成像时间长限制了其应用。Katsevich算法的各种软硬件优化方案不断被提出,但仍然难以满足实际应用中的要求。GPU的硬件结构决定了GPU的并行特点及超强的运算性能,CUDA编程技术使GPU用于通用计算领域更加便捷,两者给Katsevich算法并行加速提供了条件。在Katsevich算法实现多切片重建的过程中,反投影步骤是计算最复杂、资源消耗最大的步骤。本文在锥束覆盖理论的基础上,提出了一种基于相邻切片旋转对称性的反投影法。通过设定切片间距为射线源Z轴间距的整数倍,构造出相邻切片旋转对称性。利用该性质,对当前切片上待重建点的对应的投影数据信息进行平移和旋转操作,可以得出新切片对应的投影数据贡献度。投影数据贡献范围使用判决边界进行判断,继而使用旋转矩阵加速判决边界计算。相比于PI线法,时间和空间复杂度都有降低。使用Matlab进行了相关仿真分析并成功重建出清晰的切片影像,验证了算法的有效性。本文结合重建算法和GPU特点,给出了优化后的Katsevich算法的滤波和反投影子模块的并行加速方案。在微分求导过程中合理组织线程块,充分提高SP利用率,减少内存访问冲突;在插值模块中,使用常量内存存储插值变换模板及插值系数,常量内存的使用也加快了访问速度;在离散希尔伯特模块,使用傅里叶变换对处理输入实信号的卷积,减少计算量,使用CUFFT函数库实现FFT变换对;在反投影模块,初始化判决边界数组作为模板存放在全局内存空间,待重建切片根据距初始切片的位置使用旋转矩阵更新对应的判决边界数组,并且一个thread负责一列体素的重建。在保证重建质量的前提下,相比于CPU重建,GPU重建实现了130倍左右的加速比。相比于PI线法反投影重建2563规模切片,本文提出的基于旋转对称性的反投影不仅重建时间减少0.4秒,内存消耗仅仅是三分之一。
[Abstract]:Spiral CBCT is widely used in clinical detection, ray flaw detection and other industries. With the rapid development of electronic information, sensors and application requirements, the reconstruction quality, imaging speed and resource consumption are higher requirements. As an accurate reconstruction algorithm, Katsevich algorithm can guarantee the imaging quality. However, due to the complexity of calculation and the long imaging time, various software and hardware optimization schemes of its application. Katsevich algorithm have been proposed continuously. However, it is still difficult to meet the requirements of practical applications. The hardware structure of .GPU determines the parallelism of GPU and the super performance of CUDA programming technology, which makes it more convenient for GPU to be used in the field of general computing. Both of them provide the conditions for the parallel acceleration of Katsevich algorithm. In the process of multi-slice reconstruction by Katsevich algorithm, the step of backprojection is the most complicated and the most resource consuming step. Based on the theory of cone beam covering, this paper presents an inverse projection method based on rotational symmetry of adjacent slices. The rotation symmetry of adjacent slices is constructed by setting the slice spacing as integral times of the Z axis spacing of the ray source. By using this property, the corresponding projection data information of the current slice to be reconstructed is translated and rotated, and the contribution of the projection data corresponding to the new slice can be obtained. The contribution range of the projection data is judged by the decision boundary, and then the calculation of the decision boundary is accelerated by the rotation matrix. Compared with Pi line method, the time and space complexity are reduced. The correlation simulation analysis with Matlab is carried out and the clear slice image is reconstructed successfully, which verifies the validity of the algorithm. Combined with the characteristics of reconstruction algorithm and GPU, this paper presents a parallel acceleration scheme for filtering and backprojection submodules of the optimized Katsevich algorithm. In the process of differential derivation, the thread blocks are reasonably organized, the SP utilization ratio is improved fully, and the memory access conflict is reduced. In the interpolation module, the constant memory is used to store the interpolation transform template and interpolation coefficient, and the constant memory usage also speeds up the access speed. In the discrete Hilbert module, Fourier transform is used to deal with the convolution of input real signal, the computation is reduced, and the CUFFT function library is used to realize the FFT transform pair. The initialized decision boundary array is stored in the global memory space as a template. The slice to be reconstructed updates the corresponding decision boundary array with the rotation matrix according to the position from the initial slice, and a thread is responsible for the reconstruction of a list of voxels. On the premise of guaranteeing the reconstruction quality, the acceleration ratio of CPU reconstruction is about 130times higher than that of CPU reconstruction. Compared with the Pi line method, the proposed backprojection based on rotation symmetry not only reduces the reconstruction time by 0.4 seconds, but also consumes only 1/3 memory.
【学位授予单位】：哈尔滨工业大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：TP391.41
本文编号：2198075
[Abstract]:Spiral CBCT is widely used in clinical detection, ray flaw detection and other industries. With the rapid development of electronic information, sensors and application requirements, the reconstruction quality, imaging speed and resource consumption are higher requirements. As an accurate reconstruction algorithm, Katsevich algorithm can guarantee the imaging quality. However, due to the complexity of calculation and the long imaging time, various software and hardware optimization schemes of its application. Katsevich algorithm have been proposed continuously. However, it is still difficult to meet the requirements of practical applications. The hardware structure of .GPU determines the parallelism of GPU and the super performance of CUDA programming technology, which makes it more convenient for GPU to be used in the field of general computing. Both of them provide the conditions for the parallel acceleration of Katsevich algorithm. In the process of multi-slice reconstruction by Katsevich algorithm, the step of backprojection is the most complicated and the most resource consuming step. Based on the theory of cone beam covering, this paper presents an inverse projection method based on rotational symmetry of adjacent slices. The rotation symmetry of adjacent slices is constructed by setting the slice spacing as integral times of the Z axis spacing of the ray source. By using this property, the corresponding projection data information of the current slice to be reconstructed is translated and rotated, and the contribution of the projection data corresponding to the new slice can be obtained. The contribution range of the projection data is judged by the decision boundary, and then the calculation of the decision boundary is accelerated by the rotation matrix. Compared with Pi line method, the time and space complexity are reduced. The correlation simulation analysis with Matlab is carried out and the clear slice image is reconstructed successfully, which verifies the validity of the algorithm. Combined with the characteristics of reconstruction algorithm and GPU, this paper presents a parallel acceleration scheme for filtering and backprojection submodules of the optimized Katsevich algorithm. In the process of differential derivation, the thread blocks are reasonably organized, the SP utilization ratio is improved fully, and the memory access conflict is reduced. In the interpolation module, the constant memory is used to store the interpolation transform template and interpolation coefficient, and the constant memory usage also speeds up the access speed. In the discrete Hilbert module, Fourier transform is used to deal with the convolution of input real signal, the computation is reduced, and the CUFFT function library is used to realize the FFT transform pair. The initialized decision boundary array is stored in the global memory space as a template. The slice to be reconstructed updates the corresponding decision boundary array with the rotation matrix according to the position from the initial slice, and a thread is responsible for the reconstruction of a list of voxels. On the premise of guaranteeing the reconstruction quality, the acceleration ratio of CPU reconstruction is about 130times higher than that of CPU reconstruction. Compared with the Pi line method, the proposed backprojection based on rotation symmetry not only reduces the reconstruction time by 0.4 seconds, but also consumes only 1/3 memory.
【学位授予单位】：哈尔滨工业大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：TP391.41
【参考文献】
相关期刊论文 前3条
1 郑海亮;李兴东;王哲;魏存峰;常彤;;高分辨锥束CT并行重建算法在基于NVDIA GPU显卡计算平台上的实现[J];CT理论与应用研究;2014年05期
2 吴笃蕃;李亮;张丽;陈志强;康克军;;一种实用的基于GPU的Katsevich算法的实现(英文)[J];中国体视学与图像分析;2011年04期
3 郭晓虎,杨建生,孔强,周铁,姜明;锥束覆盖方法的并行实现及性能分析[J];中国体视学与图像分析;2005年03期
相关硕士学位论文 前2条
1 李欢;球面探测器CT图像重建算法研究及其在GPU上的实现[D];哈尔滨工业大学;2013年
2 马俊峰;三维锥束CT图像重建加速技术研究[D];山东大学;2011年
，本文编号：2198075
本文链接：https://www.wllwen.com/kejilunwen/ruanjiangongchenglunwen/2198075.html
