Performance Comparison of 3D Sinc Interpolation for fMRI Motion Correction by Language of Implementation and Hardware Platform
Abstract
Substantial effort is devoted to improving neuroimaging data processing; this effort however, is typically from the algorithmic perspective only. I demonstrate that substantive running time performance improvements to neuroscientific data processing algorithms can be realized by considering their implementation. Focusing specifically on 3D sinc interpolation, an algorithm used for processing functional magnetic resonance imaging (fMRI) data, I compare the performance of Python, C and OpenCL implementations of this algorithm across multiple hardware platforms. I also benchmark the performance of a novel implementation of 3D sinc interpolation on a field programmable gate array (FPGA). Together, these comparisons demonstrate that the performance of a neuroimaging data processing algorithm is significantly impacted by its implementation. I also present a case study demonstrating the practical benefits of improving a neuroscientific data processing algorithm's implementation, then conclude by addressing threats to the validity of the study and discussing future directions.