Smoothing and cluster thresholding for cortical surface-based group analysis of fMRI data

Cortical surface-based analysis of fMRI data has proven to be a useful method with several advantages over 3-dimensional volumetric analyses. Many of the statistical methods used in 3D analyses can be adapted for use with surface-based analyses. Operating within the framework of the FreeSurfer software package, we have implemented a surface-based version of the cluster size exclusion method used for multiple comparisons correction. Furthermore, we have a developed a new method for generating regions of interest on the cortical surface using a sliding threshold of cluster exclusion followed by cluster growth. Cluster size limits for multiple probability thresholds were estimated using random field theory and validated with Monte Carlo simulation. A prerequisite of RFT or cluster size simulation is an estimate of the smoothness of the data. In order to estimate the intrinsic smoothness of group analysis statistics, independent of true activations, we conducted a group analysis of simulated noise data sets. Because smoothing on a cortical surface mesh is typically implemented using an iterative method, rather than directly applying a Gaussian blurring kernel, it is also necessary to determine the width of the equivalent Gaussian blurring kernel as a function of smoothing steps. Iterative smoothing has previously been modeled as continuous heat diffusion, providing a theoretical basis for predicting the equivalent kernel width, but the predictions of the model were not empirically tested. We generated an empirical heat diffusion kernel width function by performing surface-based smoothing simulations and found a large disparity between the expected and actual kernel widths.