The Influence of Spatial Smoothing Kernel Size on ICA Model Order and Spatial Maps of Intrinsic Connectivity Networks

Abstract

Earlier studies indicate that fMRI preprocessing methods can affect the properties of the brain intrinsic connectivity networks (ICNs). Previously, we showed that spatial smoothing, a standard preprocessing step, would influence time-varying whole-brain network connectivity patterns and meta-states metrics. Here, we study the influence of spatial smoothing on the dimensionality of the fMRI data and ICN spatial maps. To this end, we collected resting-state fMRI data of healthy subjects using a 3.0 T MRI scanner. During preprocessing, we applied various levels of spatial smoothing to the data with an isotropic Gaussian kernel with full width at half maximum (FWHM) sizes 0 to 12 mm with a step of 2 mm and calculated ICA model order to estimate the number of informative components. We examined the significant changes in the spatial maps of the data that were preprocessed with 4, 8, and 12 mm smoothing kernels pairwise using a paired $t$-test with a false discovery rate correction. Results revealed that the level of spatial smoothing clearly impacts the network dimensionality, intensities of spatial maps, and peak voxel location. Using minimum description length (MDL) criteria, dimensionality generally decreased as smoothing kernel size increased. In contrast, entropy-rate based order selection indicated a general increase in model order as smoothing kernel size increased. Intensities of spatial maps, which are associated with the cohesiveness and connectivity inside the network, decreased in most ICNs, including the default-mode and salience networks as the smoothing kernel size decreased. Our findings provide a preliminary insight into the effects of spatial smoothing on ICA model order and spatial maps. Larger samples are needed to further investigate these effects.