Statistical image reconstruction for cone-beam X-ray CT on a shared memory computation platform

Statistical reconstruction methods offer possibilities to improve image quality as compared to analytical methods, but current reconstruction times prohibit routine clinical applications. To reduce reconstruction times we have parallelized a statistical reconstruction algorithm for cone-beam X-ray CT, the ordered subset convex algorithm (OSC), and evaluated it on a shared memory computer. Two different parallelization strategies were developed: one that employs parallelism by computing the work for all projections within a subset in parallel, and one that divides the total volume into parts and processes the work for each subvolume in parallel. Both methods are used to reconstruct a three dimensional mathematical phantom on two different grid densities. The reconstructed images are binary identical to the result of the serial (non-parallelized) algorithm. The speed-up factor equals approximately 30 when using 32 to 40 processors, and scales almost linearly with the number of cpus for both methods. The huge reduction in computation time allows to apply statistical reconstruction to clinically relevant studies for the first time.

I. Introduction

Statistical image reconstruction (SR) has proven to be a very valuable tool in medical imaging, for both emission and transmission computed tomography (CT), see e.g. [1] – [5]. However, a known disadvantage of using SR for X-ray CT is its long reconstruction time. One way to reduce reconstruction time is to parallelize the algorithm, a well known approach in tomographic imaging. For example in Positron Emission Tomography [6] – [8] and Single Photon Emission Computed Tomography [9] – [12] as well as other image modalities [13], [14] parallelization has proven to yield significant reductions of reconstruction time. However, in CT relatively little work in this direction has been done, see e.g. [15], [16], partly because SR is not often used for image reconstruction in CT yet.