PMC

Time per iteration (in seconds) as a function of number of subsets.

Line profiles of the image background for SART, SIRT, and the original Barbara image.

Optimal relaxation factor λ as a function of number of subsets for the optimal λ-selection scheme.

CC vs. wall clock time for the linear λ-selection scheme. We observe that OS-SIRT achieves the reconstruction it in the smallest amount of time, within our GPU-accelerated framework.

CC vs. wall clock time for the optimized λ–selection scheme. We observe that there is now a clear ordering in terms of reconstruction time from small subsets to larger ones.

Reconstructions with the optimized λ-selection scheme obtained with various subset sizes for a fixed CC=0.95 and 180 projections in an angular range of 180°. We observe that now the number of subsets is directly related to wall clock computation time, with SART being the fastest.

Reconstructions obtained with the linear λ-selection schedule for various subset sizes for a fixed CC=0.95 and 180 projections in an angular range of 180°. We observe that OS SIRT with 10 subsets of 18 projections each reaches this fixed CC value 12% faster than SIRT and 50% faster than SART.

Reconstruction of the baby head from a limited set of 64 projections in an angular range of 180° and the optimized λ–selection scheme shown above with various subset sizes for a fixed R-factor = 0.007. We observe that OS-SIRT with 16 subsets of 4 projections each is the fastest in this case. It reaches this set R-factor value 84% faster than SIRT and 56% faster than SART.

Reconstructed baby head using high-quality simulated projection data of the volume labeled ‘Original’. Results were obtained with the linear λ–selection scheme with various subset sizes for a fixed R-factor = 0.007 and 180 projections in an angular range of 180°. OS-SIRT with 10 subsets of 18 projections each reaches this set R-factor value 26% faster than SIRT and 86% faster than SART.

Reconstruction results of the baby head obtained with the optimized λ–selection scheme with various subset sizes for a fixed R-factor = 0.007 and 180 projections in an angular range of 180°. Using the λ-selections shown above, OS-SIRT with 20 subsets of 9 projections each is the fastest in this case. It reaches this set R-factor value 91% faster than SIRT and 72% faster than SART. We also note that the time is about 7 times faster than with the linear selection scheme since the optimal λ-factor is higher.