Performance of Image Matching in the Computational Anatomy Gateway: CPU and GPU Implementations in OpenCL

The Computational Anatomy Gateway is a software as a service tool that provides tools for analysis of structural MRI to the neuroimaging community by calculating diffeomorphic mappings between a user's data and well characterized atlas images. These tools include automatic parcellation of brain images into labeled regions, described by dense 3D arrays; and shape analysis of regions described by triangulated surfaces, for hypothesis testing in specific populations. We have developed mapping techniques that combine the benefits of working with triangulated surfaces with those of working with dense images, and have been working toward uniting these two tools: to automatically perform shape analysis on each segmented subcortical structure simultaneously.

In this work we investigate the performance of our algorithm across a wide range of input data, examining the effect of number of voxels in 3D images, number of vertices in triangulated surfaces, and number of structures being mapped onto simultaneously. Further, we investigate the performance of our OpenCL code implemented in two different environments: the Intel OpenCL environment on a CPU, and the CUDA OpenCL environment on a GPU.

We identify a range of inputs, generally smaller datasets, for which the CPU out performs the GPU. Finally we show the feasibility of mapping onto all the human gray matter subcortical structures simultaneously, and discuss our strategy for extending to higher resolution images and more labeled structures in mouse brain imaging at the micrometer scale.