Affiliations: School of Computer Science and Engineering, Seoul
National University, 599 Gwanak-ro, Gwanak-gu, Seoul, Republic of Korea | Department of Radiation Oncology, Stanford University,
Stanford, CA, USA
Note: [] Corresponding author: Ho Lee, Department of Radiation Oncology,
Stanford University, Stanford, CA, USA. Tel.: +1 650 721 5475; Fax: +1 650 498
4015; E-mail: [email protected]
Abstract: This paper presents a fast hybrid CPU- and GPU-based CT
reconstruction algorithm to reduce the amount of back-projection operation
using air skipping involving polygon clipping. The algorithm easily and rapidly
selects air areas that have significantly higher contrast in each projection
image by applying K-means clustering method on CPU, and then generates boundary
tables for verifying valid region using segmented air areas. Based on these
boundary tables of each projection image, clipped polygon that indicates active
region when back-projection operation is performed on GPU is determined on each
volume slice. This polygon clipping process makes it possible to use smaller
number of voxels to be back-projected, which leads to a faster GPU-based
reconstruction method. This approach has been applied to a clinical data set
and Shepp-Logan phantom data sets having various ratio of air region for
quantitative and qualitative comparison and analysis of our and conventional
GPU-based reconstruction methods. The algorithm has been proved to reduce
computational time to half without losing any diagnostic information, compared
to conventional GPU-based approaches.
