A "brick" caching scheme for 3D medical imaging

This paper presents a "brick" caching scheme for a broad range of 3D medical imaging algorithms, which achieves a cache hit ratio of 100% and significantly reduces the total data-flow from the system memory to the caches, thus reducing the memory bandwidth requirement of the system. With data indexing and duplication, this scheme supports multiple parallel memory access required in most 3D imaging algorithms. Brick prefetching strategy fetches data into the cache before they are used, therefore preventing the stall of the processing pipeline. The applications of the caching scheme in FDK cone beam reconstruction and 3D image registration were presented as examples. This caching scheme will be used in our FPGA-based 3D medical imaging coprocessor and it can also be applied to other FPGA-based 3D imaging accelerators.