Modern GPU-based Optimization of Katsevich Image Reconstruction Algorithm

Hong-ying Zhang?, Xiu-yu Sun, Yi-gang Sun

College of Aeronautical Automation, Civil Aviation University of China, Tianjin, 300300, China

ABSTRACT

Katsevich’s algorithm is the first theoretically exact spiral cone-beam reconstruction algorithm in computed tomography (CT). However, compared with traditional approximate methods, it still requires much more complicated calculations. Fortunately, the development of graphics processing unit (GPU) technologies provides a parallel accelerating approach to resolve the problem of heavy computation burden. Therefore, a GPU-based parallel implementation of Katsevich’s algorithm is proposed in this paper. We firstly optimize the memory accesses in differentiating and remapping processes, and then simultaneously perform Hilbert transform along k-lines. Finally, multi-viewed operations are adopted in the back-projection process to reduce memory accesses, and optimization between pitches is used to reduce calculation of projections’ preprocessing. Several experiments were performed to validate the proposed algorithm and results indicate that our parallel algorithm significantly reduces the reconstruction time, achieving high speedup and efficiency.

Keywords Cone-beam reconstruction, GPU-based optimization, Katsevich’s algorithm