Adaptive spatial smoothing of fMRI images

Abstract

It is common practice to spatially smooth fMRI data prior to statistical analysis and a number of different smoothing techniques have been proposed (e.g., Gaussian kernel filters, wavelets, and prolate spheroidal wave functions). A common theme in all these methods is that the extent of smoothing is chosen independently of the data, and is assumed to be equal across the image. This can lead to problems, as the size and shape of activated regions may vary across the brain, leading to situations where certain regions are under-smoothed, while others are over-smoothed. This paper introduces a novel approach towards spatially smoothing fMRI data based on the use of nonstationary spatial Gaussian Markov random fields (Yue and Speckman, 2009). Our method not only allows the amount of smoothing to vary across the brain depending on the spatial extent of activation, but also enables researchers to study how the extent of activation changes over time. The benefit of the suggested approach is demonstrated by a series of simulation studies and through an application to experimental data.

LISA improves statistical analysis for fMRI

“…A large range of adaptive filters exist that might be used in this context 44–48 . In the experiments reported below, we have used the bilateral filter because it is widely used and well known 49–51 , see also Supplementary Figure 25.…”

A Hitchhiker's Guide to Functional Magnetic Resonance Imaging

“…Smoothing simultaneously increases the SNR and the validity of the statistical tests (from random field theory) by providing a better fit to expected assumptions while reducing the anatomical differences. On the other hand, smoothing reduces the effective spatial resolution, may displace activation peaks (Reimold et al, 2006) and extinguish small but meaningful local activations depending on the filter parameters chosen (Yue et al, 2010; Poldrack et al, 2011; Sacchet and Knutson, 2013). The standard spatial smoothing procedure consists of convolving the fMRI signal with a Gaussian function of a specific width (as, spatially, the BOLD signal is expected to follow a Gaussian distribution).…”

“…The typical smoothing values used range between 5 and 10 mm for group analyses (Beckmann and Smith, 2004; Mikl et al, 2008; Poldrack et al, 2011). Alternative approaches to smoothing are the use of varying kernel widths (Worsley et al, 1996), adaptive smoothing (Yue et al, 2010; Bartés-Serrallonga et al, 2015), wavelet transforms (Van De Ville et al, 2007), and prolate spheroidal wave functions (Lindquist et al, 2006). Despite its common use, care must be taken when performing smoothing due to its effects on the final results (Geissler et al, 2005; Molloy et al, 2014), its interaction with motion correction (Scheinost et al, 2014) and impact upon analyses which are sensitive to the activation of individual voxels (such as ROI-to-ROI analysis, Regional Homogeneity and Multi-voxel Pattern Analysis).…”

Using fMRI non-local means denoising to uncover activation in sub-cortical structures at 1.5 T for guided HARDI tractography

“…The most common technique to acheive this goal is the standard global Gaussian smoothing (Cox, 1996; Friston et al, 2000b; Smith et al, 2004; Strother, 2006). This approach replaces a voxel intensity value by a distance-weighted average of its surrounding neighbors such as a full-width half maximum (FWHM) of 5 mm, though this can vary across studies (Wink and Roerdink, 2004; Yue et al, 2010). Often, the FWHM is chosen independently of the data and set equal across the image.…”

“…Often, the FWHM is chosen independently of the data and set equal across the image. This can lead to problems, as the size and shape of activated regions may vary across the brain, leading to situations where certain regions are under-smoothed, while others are over-smoothed, potentially leading to false positives (Yue et al, 2010). …”