A GRAPHICAL PROCESSING UNIT-BASED PARALLEL IMPLEMENTATION OF MULTIPLICATIVE ALGEBRAIC RECONSTRUCTION TECHNIQUE ALGORITHM FOR LIMITED VIEW TOMOGRAPHY

This article proposes an efficient two-dimensional (2D) pixel-driven multiplicative algebraic reconstruction technique (PdMART) on a general purpose graphical processing unit (GPU), Nvidia graphics card GTX-275. It has been tested on numerical data and also on real data that have been obtained from the micro-computed tomography scanner installed at University of Manchester. We have used real data having 90 projections and 256 rays in each projection to test the algorithm. The real data has been obtained by scanning the graphite core object of size 30mmx30mm. It has been found that GPU can help PdMART to generate the weight matrix for the 256x256 pixel grid within a second which is very fast compared to any sequential machine. Experimental results reveal that PdMART on GPU is computationally inexpensive. Preliminary results also indicate much better performance (as compared to popular Fourier methods) for cases of limited-view projection data as is the case for the upcoming laminographic tomography machines.