Design of an FPGA-based computing platform for real-time three-dimensional medical imaging

Real-time 3D medical imaging requires very high computational capability that is beyond most of the general computing platforms. Although application specific integrated circuits (ASIC) can provide solutions for a particular algorithm, they are too expensive to develop and most of them are not flexible enough to adapt to the evolution of existing algorithms or the emergence of new problems. FPGA-based reconfigurable architectures combined with general-purpose processors exhibit a good tradeoff in performance and flexibility, and are affordable for practical applications. To address the problems in designing such a system, including long designing and testing time, complex data manipulation and high performance requirement etc., we designed a new computing platform to accelerate a broad range of local operation-based 3D medical imaging algorithms. This platform is composed of a new data caching scheme, called brick caching scheme and a reconfigurable System-on-Chip (SoC) architecture targeted to Xilinx Virtex-II Pro FPGAs. The brick caching scheme exploits spatial locality of reference in three dimensions with 3D block caching; it enables data prefetching by obtaining input data block information through input-output space mapping; it also supports multiple data accesses with data duplication. An intelligent data caching system is built around a PowerPC processor core in the SoC architecture to support the brick caching scheme. A multiple pipeline execution unit that is reconfigurable to different algorithms is designed to perform vectorized computation. Two algorithms are implemented and tested on this platform, one is the FDK cone-beam CT reconstruction algorithm and the other is the mutual information-based 3D registration algorithm. Our simulation results demonstrate that a speed-up of about 30 can be achieved for both of the algorithms.