High resolution iterative CT reconstruction using graphics hardware

Particular applications of computed tomography require high slice resolutions. The fastest iterative implementations on graphics cards use 3-D textures to exploit hardware-accelerated trilinear interpolation. However, the size of 3-D textures is subject to technical limitations, which makes them inapplicable here. Alternatively a 2-D texture array can be used instead of the 3-D texture as in early graphics implementations. The additional memory synchronizations cause a significant loss of performance. We utilize new features of the recently released CUDA 2.2 framework to improve the performance of the Simultaneous Algebraic Reconstruction Technique (SART). In this paper we present an enhanced version of our efficient implementation of the most time-consuming parts of the iterative reconstruction algorithm: forward- and back-projection. We explain the required strategy to adapt the algorithm for the CUDA 2.2 features, in particular the usage of 2-D texture lookups from pitchlinear memory. Finally, we compare the result to our previous ones with respect to both reconstruction speed and technical limitations. The proposed strategy is a new balance between performance and limitations in resolution.