Searchlight Goes GPU - Fast Multi-Voxel Pattern Analysis of fMRI Data

Graphics processing units (GPUs) are used today in a wide range of applications, mainly because they can dramatically accelerate parallel computing, are affordable and power efficient. For fMRI, GPUs have recently been used to speedup non-parametric statistical methods, which can be more reliable than parametric methods used by most software packages. Here we demonstrate the utility of GPUs for multivariate fMRI analysis. The searchlight algorithm is a popular choice for locally-multivariate decoding of fMRI data. For each voxel, a classifier is trained to discriminate between different brain states, by using voxels within a neighborhood search volume. A performance measure, typically the classification accuracy, is then saved in the center voxel. A substantial drawback of the searchlight is that it is computationally demanding. This is especially true for large searchlight spheres, non-linear classifiers, cross validation schemes and statistical permutation testing. Here we therefore present a GPU implementation of the searchlight algorithm, which is over 8000 times faster than a simple Matlab implementation and about 21 times faster than a parallelized Matlab implementation.