The original filtered backprojection describes a loop through each pixel in the image, locating the nearest rays for the corresponding pixel. The modified version however uses each ray as the center of attention. These are traced through the image, adding to the pixels that are intersected. Due to this modification, the image can be reconstructed entirely by the use of integers. To further optimize, the image is divided in sixteen squares and each square is reconstructed by a synergistic processing element (SPE). For the cases where the numbers of SPEs are less than sixteen, a static division of the squares is implemented. The implementation is tested on a Playstation 3 with 256MB memory and six available SPEs. The system is using Fedora Core 7 as OS. Tested on 1, 2, 3, 4 and 6 SPEs, the result is a parallel algorithm with near linear speedup. The results of the parallel algorithm is compared to a sequential implementation, executed on an Intel Core 2 Duo processor with 2GB memory and also running Fedora Core 7. This test shows that using a single SPE is faster than the commodity processor, despite the overhead in terms of memory transfers and bookkeeping. This is caused by the vector-units residing in each SPE, enabling a boost in performance for certain types of algorithms [1].

Since the backprojection algorithm is modified, a comparison is made to an implementation of the original backprojection, performed on the same dataset. The test shows that the modified version is competitive in terms of quality, however, since the test is performed on synthetic data, a final conclusion cannot be drawn.

1

IBM, "Cell Broadband Engine Programming Handbook", 1.1 edition, 2007.

2

A.C. Kak, M. Slaney, "Principles of Computerized Tomographic Imaging", Society of Industrial and Applied Mathematics, 2001. doi:10.1118/1.1455742

purchase the full-text of this paper (price £20)

go to the previous paper
go to the next paper
return to the table of contents
return to the book description
purchase this book (price £72 +P&P)