Algorithm Architecture Adequacy for High Speed 3D Tomography
Abstract
Backprojection is a computational costly step in tomography image reconstruction such as Positron Emission Tomography (PET). In this purpose, this paper presents a Pipelined, Prefetch and Parallelized Architecture for PET backprojection (3PA-PET). The main strength of this architecture comes from its original memory access strategy, masking the high memory latency of the external memory. The 3PAPET architecture is implemented on a System on Programmable Chip (SopC). Time performances are compared with a desktop PC, a workstation and a GPU. We prove that the exploitation of the intrinsic temporal and spatial locality by the 3D Predictive and Adaptative (3D-AP) memory cache succeeds to run efficiently several pipelines of backprojection : each reaches a computational throughput close to 1 operation per cycle.
Origin : Files produced by the author(s)