ROTATION BASED ALGORITHM FOR PARALLELIZING OS-SART FOR CT ON HOMOGENOUS MULTICORE ARCHITECTURE

摘要

Iterative reconstruction techniques are safe approaches forrncomputed tomography (CT). However, the long processingrntime of them prevent their adoption in clinical CT machines.rnRecently, with the growing increase in computationalrnpower due to multiple cores on a chip, iterative techniquesrnare no longer deemed unsolvable. In this paper,rnwe investigate one category of iterative techniques whichrnstem from algebraic reconstruction technique, includingrnART (algebraic reconstruction technique), SIRT (simultaneousrniterative reconstruction technique), SART (simultaneousrnalgebraic reconstruction technique), and orderedrnsubset version of SART (OS-SART). We exploit the parallelismrnin OS-SART and its feasibility to parallelization onrnhomogenous shared memory multicore architectures. OSSARTrnis memory bound due to the frequent access to arnlarge weighting factor matrix. Therefore, to reduce memoryrnlatency in OS-SART, we parallelize OS-SART algorithmrnby incorporating rotation-based projector and backprojectorrnwhich enables the computation of weighting factorsrnat runtime on a need basis. The results show that ourrnparallel algorithm is scalable on homogeneous multicorernarchitecture in terms of both core number and problemsize.rnFor 8 cores, the speedup increases from 6.77 for imagernsize of 128x128 to 7.02 for image size of 512x512. Forrn16 cores, the speedup increases from 12.29 for image sizern128x128 to 13.18 for image size 512x512.