Optimization For CBCT Multi-slice Reconstruction Based On Katsevich Algorithm And Implementation On GPU

Helical CBCT is widely used in medical and industrial fields.With the development of technology and application requirements,higher requirements are put forward about imaging quality,reconstruction speed and resource consumption.As an exact reconstruction algorithm,the Katsevich algorithm can guarantee the image quality.But the complex calculation with long reconstruction time prohibits its application.There are many optimization strategy of Katsevich algorithm both in hardware and software,it's still hard to meet the practical application requirements.GPU has superior computing performance as its multi-core hardware architecture,CUDA programming technology makes GPU development more convinient.So using GPU accelerate the algorithm is a good choice.The back-projection process is the most complex step.The dissertation proposes a novel back-projection method based on the rotational symmetry of neighboring slices by seting slices' stepsize be the multiple of x-ray sources' stepsize.Taking advantage of this feature,the contribution of projection data to new slices can been updated quickly.And decision boundary is put forward to calculate contribution field.Furthermore,a rotation matrix is used to accelerate the decision boundary computing.The novel method is analyzed and verified on maltab.The clear reconstrution image is obtained.The dissertation designs the parallel acceleration strategies for each sub-module of the optimized Katsevich algorithm.In the differential derivative module,the threads and blocks are organized properly to improve SP utilization and reduce the global memory access conflict.In the interpolation module,the constant memory is used to store the interpolation transform and coefficient template.As for the discrete Hilbert transform,the FFT is used instead of convoluting directly.And CUFFT function library is utilized to excute the FFT transform.In the back-projection module,the decision boundary array is initialized firstly,and one thread reconstruct one column voxel.Experiments show that GPU reconstruction can achieve 130 times speed ration with acceptable reconstructed image quality.Compared to PI-line method,reconstructing the 2563 slices can reduce 0.4 s,while the memroy consumption can reduce 60%.