Efficient ray tracing on 3D regular grids for fast generation of digitally reconstructed radiographs in iterative tomographic reconstruction techniques

Cone beam projection is an essential and particularly time consuming part of any iterative tomographic reconstruction algorithm. On current graphics hardware especially the amount and pattern of memory accesses is a limiting factor when read-only textures cannot be used. With the final objective of accelerating iterative reconstruction techniques, a non-oversampling Joseph-like raytracing projection algorithm for three dimensions featuring both a branchless sampling loop and a cache friendly memory access pattern is presented. An interpretation of the employed interpolation scheme is given with respect to the effective beam and voxel models implied. The method is further compared to existing techniques, and the modifications required to implement further voxel and beam shape models are outlined. Both memory access rates and total run time are benchmarked on a current consumer grade graphics processing unit and explicitly compared to the performance of a classic Digital Differential Analyzer (DDA) algorithm. The presented raytracer achieves memory access rates of 292 GB/s in read-and-write memory and 502 GB/s in read-only texture memory. It outperforms the DDA in terms of total run time by a factor of up to five and achives 170 to 300 projections of a 512 voxel volume per second.