A fast multi-core implementation of the 3DRP algorithm

Objectives The 3D re-projection (3DRP) algorithm is a computationally challenging algorithm and its applicability is limited by the amount of time required to reconstruct the images. 3DRP has better contrast recovery than 2-D reconstruction algorithms using faster rebinning methods and also performs much better in studies that involve hot regions. In this article, we propose a multi-core implementation of 3DRP to significantly reduce reconstruction time.

Methods We used Pthreads to achieve parallelism. 3DRP reconstruction involves forward projection to estimate missing projection data (reprojection), 2-D filtration of completed projection data using the Colsher filter, and finally backprojection of filtered projection at each view. While reprojection, filtration and backprojection must be performed for all views, each view can be processed independently. Hence, to improve speed of 3-D reconstruction speed we utilized multiple cores by equally dividing the computations among cores based upon number of views. We reconstructed the image of a micro-Derenzo phatom, having histogrammed the listmode data with ring difference of 79 and span of 3 using high-resolution settings for the image sizes of 128×128×159 and 256×256×159. The reconstructions were first performed using just one core and then again using multiple cores. These reconstructions were performed on a DTx E5440 with two CPUs, each comprising four cores running at 2.83 GHz, and 32 GB of RAM.

Results We observed a speedup factor of >6 as summarized in Table 1. We also calculated the difference between the images reconstructed with single core and multiple cores, which was on the order of floating point precision.

Conclusions Our new approach results in a significant speed improvement. This improvement is not a linear function of number of processors due to some read/write hard disk activity, different RAM utilizations, and also because we only parallelize the computationally expensive tasks.