SPM/Spatial smoothing

Spatial smoothing is usually performed as a part of the preprocessing of individual brain scans.

In SPM the spatial smoothing is performed with a spatially stationary Gaussian filter where the user must specify the kernel width in mm "full width half max". Kernel widths of up to 16mm are being used in the literature.

The purpose of spatial smoothing is to cope with functional anatomical variability that is not compensated by spatial normalization ("warping") and to improve the signal to noise ratio. In other words: smoothing increases statistical power. The less one smoothes, the less likely it is to obtain significant results. Understanding the exact impact of spatial smoothing on statistical power is not trivial: smoothing not only increases the signal to noise ratio of each voxel, it also reduces the number of resolution elements ("resels") that are assumed to be independent and used for correction of multiple testing. With respect to the signal to noise ratio, it is being argued that an ideal filter kernel matches the size of the signal to detect [1]. After all, the level of smoothing that is optimal for real, measured data can only be determined empirically [2] and it may make sense to reanalyze the same data with several levels of smoothing [3].

The drawback from smoothing is a loss of spatial precision. Especially when using larger smoothing kernels (and when separately estimating the local variance for each smoothed voxel, which has been the standard in SPM, including SPM2, [4]), one should keep in mind that the loss of spatial precision is more dramatic in the resulting t-map than in the images itself due to nonlinear effects on the voxel variances [5]. In the worst case, the maximum t-value may be located in an area with no signal at all (e.g. white matter).