Parallel-beam backprojection: an FPGA implementation optimized for medical imaging

Reviewer: John A. Fulcher

Leeser et al. investigate the use of field programmable gate arrays (FPGA) as an alternative approach to both digital signal processors and application-specific integrated circuits in medical imaging. The application of interest in this paper is the filtered backprojection algorithm, which is a fundamental step in tomographic image reconstruction. Parallel-beam computerized tomographic scanning involves the use of an array of equally spaced unidirectional sources of focused x-ray beams. The radon transform maps an image into a sinogram; the filtered backprojection algorithm performs the inverse transformation. The authors found that several modifications were required to render the basic algorithm efficient for implementation in hardware form. The most important was the use of fixed-point arithmetic rather than floating-point arithmetic, for simplicity (although this does lead to higher quantization, rounding, and truncation errors). Further, a compromise bit width was chosen-too narrow means less scope for exploiting the inherent parallelism offered by FPGAs; too wide leads to unacceptably long computation times (although from the perspective of medical image precision, the wider the better). A worst-case relative error of 0.015 percent (relative, that is, to a floating-point implementation) resulted from adopting the above simplifications. The authors compare results obtained by way of software simulation (on a 1GHz Pentium personal computer with 256KB of cache) to those obtained using a Xilinx Virtex 2000E FPGA (mounted on an Annapolis Micro Systems Firebird PCI system expansion board). Computation times for 1024x1024 projection values for a 512x512 pixel image were 28 seconds (software), 3.6 seconds, 0.5 seconds, and 0.25 seconds (for sequential, 8-way parallel, and 16-way parallel FPGAs, respectively)-in other words, a speedup of two orders of magnitude. In the future, the authors plan to investigate the feasibility of applying their approach to fan-beam reconstruction, segmentation and reassembly image formation, and cone beam reconstruction. This paper will appeal to readers with an interest in medical imaging, FPGAs, or signal processing in general. Online Computing Reviews Service