Comprehensive analysis of high-performance computing methods for filtered back-projection

This paper provides an extensive runtime, accuracy, and noise analysis of Computed To-mography (CT) reconstruction algorithms using various High-Performance Computing (HPC) frameworks such as: “conventional” multi-core, multi threaded CPUs, Compute Unified Device Architecture (CUDA), and DirectX or OpenGL graphics pipeline programming. The proposed algorithms exploit various built-in hardwired features of GPUs such as rasterization and texture filtering. We compare implementations of the Filtered Back-Projection (FBP) algorithm with fan-beam geometry for all frameworks. The accuracy of the reconstruction is validated using an ACR-accredited phantom, with the raw attenuation data acquired by a clinical CT scanner. Our analysis shows that a single GPU can run a FBP reconstruction 23 time faster than a 64-core multi-threaded CPU machine for an image of 1024 X 1024. Moreover, directly programming the graphics pipeline using DirectX or OpenGL can further increases the performance compared to a CUDA implementation.