Skip to main page content
Access keys NCBI Homepage MyNCBI Homepage Main Content Main Navigation
. 2014 Jul 1;8:462.
doi: 10.3389/fnhum.2014.00462. eCollection 2014.

Deficient Approaches to Human Neuroimaging

Affiliations
Free PMC article

Deficient Approaches to Human Neuroimaging

Johannes Stelzer et al. Front Hum Neurosci. .
Free PMC article

Abstract

Functional magnetic resonance imaging (fMRI) is the workhorse of imaging-based human cognitive neuroscience. The use of fMRI is ever-increasing; within the last 4 years more fMRI studies have been published than in the previous 17 years. This large body of research has mainly focused on the functional localization of condition- or stimulus-dependent changes in the blood-oxygenation-level dependent signal. In recent years, however, many aspects of the commonly practiced analysis frameworks and methodologies have been critically reassessed. Here we summarize these critiques, providing an overview of the major conceptual and practical deficiencies in widely used brain-mapping approaches, and exemplify some of these issues by the use of imaging data and simulations. In particular, we discuss the inherent pitfalls and shortcomings of methodologies for statistical parametric mapping. Our critique emphasizes recent reports of excessively high numbers of both false positive and false negative findings in fMRI brain mapping. We outline our view regarding the broader scientific implications of these methodological considerations and briefly discuss possible solutions.

Keywords: brain mapping; cognitive neuroscience; critical neuroscience; fMRI; functional localization.

Figures

FIGURE 1
FIGURE 1
The usage of functional MRI gets increasingly popular. We depict the number of publications for each year that incorporate fMRI on human subjects. The data is based on a pubmed.org search (see appendix).
FIGURE 2
FIGURE 2
(A) Simplified basic experimental rationale underlying fMRI-based brain-mapping studies. Two (carefully chosen) experimental conditions elicit two distinct brain states. FMRI measurements are able to capture certain aspects of these brain states. The resulting images are compared statistically and this statistical difference is causally ascribed to the difference in terms of experimental factors. (B) Minimalistic preprocessing and data analysis pipelines. Preprocessing includes head-motion correction, transformation into a common coordinate system (e.g., Talairach or MNI space) and spatial smoothing. On this preprocessed data, model estimates are computed (here, we depicted the general linear model). The statistical comparison across the group usually is carried out on basis of the model estimates (e.g., contrasts) and incorporates a correction for the multiple comparisons problem.
FIGURE 3
FIGURE 3
The effects of spatial smoothing illustrated on an ultra-high field fMRI data set at a field strength of 7 Tesla and an isotropic resolution of 0.65 mm. The scanning paradigm comprised a visual checkerboard stimulation (see appendix) and the analysis was based on a simple general linear model. In the left column, the results are displayed in terms of a t-statistic (i.e., significance of activations). In the right column, the corresponding effect sizes are shown (i.e., amplitude of activations). We depicted the original results and four levels of spatial smoothing. When smoothing is omitted, fine-grained activation patterns are visible on the cortical surface (i.e., within gray matter regions). While spatial smoothing increases the statistical significance of the results, both the effect size and spatial accuracy of the results are drastically reduced. Noteworthy, the intrinsic SNR would be increased if a larger voxel size was used or more repetitions were carried out. As result, a larger number of voxels would be labeled active.
FIGURE 4
FIGURE 4
After the application of spatial smoothing, the signal of a given voxel is a mix between its original signal and the weighted average of its surroundings. Smoothing implies that for any given voxel (A) the signal is “washed” into its neighborhood and (B) that signal of its surroundings is washed into this voxel. (C) We display the ratio of this mix, that is, how much of the original signal remains in the voxel versus how much signal of the neighborhood is washed in. We computed this mix for different smoothing kernel sizes (given in the dimension of voxels, for details see appendix). The blue bars represent the original (internal) signal of a voxel; the red bars illustrate the fraction of the (external) signal that stems from the neighborhood of the voxel. It is visible that already for relatively small smoothing kernels (e.g., two voxels, corresponding to a FWHM of 6 mm given 3 mm voxels), more than 90% of the signal does not correspond any more to the original signal at any given voxel but stems from the voxel’s neighborhood. This implies a severe loss of spatial precision for functional localization.
FIGURE 5
FIGURE 5
The effects of combined spatial smoothing and cluster-based statistics are additive. We used simulations to depict the minimum cluster size which is determined as significant for various levels of smoothness (see appendix). Clusters of smaller extent than this minimum size fail to reach significance and are effectively sieved out. For instance, if an overall smoothness of 4 voxels is assumed in the underlying images, only clusters that are larger than about 80 voxels are taken into consideration (i.e., are corresponding to an uncorrected cluster p-value < 0.05). It should be noted, however, that this uncorrected cluster p-value is still subject to a multiple comparisons correction; hence depending on the number of overall clusters the final minimum cluster size may actually be considerably larger than the value we depict here.
FIGURE 6
FIGURE 6
Thought experiment considering the activation patterns for three subjects (A,B,C). The individual (ground truth) activations of the three subjects are displayed in the colors red, green, and blue. Critically, we assume the true activations to be variable across the subjects. If standard group statistical procedures then are applied on this scenario, only the effective overlap of the subjects is revealed (D). We display this effective overlap in an orange “blob-like” tone in (D), for the sake of illustration we marked the overlap also in the individual subject patterns (A,B,C) using white dots. Under the assumption of high inter-individual variance, this illustration shows the fallacy of spatial group statistics: for none of the subjects the overlap regions were sufficient for representing the brain state, as each subject relied on the involvement of further regions.
FIGURE 7
FIGURE 7
Test and retest study, where two participants (JK and MM) were scanned on two separate sessions, held months apart. The arrows in between the participants and sessions denote the cross-correlation between the (uncorrected) patterns of brain activity, thus indicating their degree of similarity. While the spatial patterns remain relatively similar within a subject (across both sessions), the similarity across subjects is comparably low. Reprint with permission from Miller et al. (2012).
FIGURE 8
FIGURE 8
Data simulation investigating the effects of a heterogeneous subject group (in terms of responders and non-responders). We computed group statistics (t-based) for one single location using a group of 20 “virtual” subjects (see appendix). This group consisted of two subgroups, firstly responders and secondly non-responders. The former were sampled from an effect distribution (normal distribution with an offset) and the latter from a null distribution (normal distribution). We varied the composition of the group (i.e., the number of responders versus the number of non-responders) and also the size of the effect (the offset). For each level (20 levels of group composition and six levels of effect sizes) we repeatedly computed a t-based one-sample random-effects analysis, as used in common group level inference. We displayed the resulting statistical significance levels in different colors: green for mild significance (0.01 < p < 0.05), blue for moderate significance (0.001 < p < 0.01) and red for high significance (p < 0.001).

Comment in

Similar articles

See all similar articles

Cited by 22 articles

See all "Cited by" articles

References

    1. Bazin P.-L., Weiss M., Dinse J., Schäfer A., Trampel R., Turner R. (2014). A computational framework for ultra-high resolution cortical segmentation at 7 Tesla. Neuroimage 93 201–209 10.1016/j.neuroimage.2013.03.077 - DOI - PubMed
    1. Benjamini Y., Heller R. (2007). False discovery rates for spatial signals. J. Am. Stat. Assoc. 102 1272–1281 10.1198/016214507000000941 - DOI
    1. Bianciardi M., Fukunaga M., van Gelderen P., Horovitz S. G., de Zwart J. A., Shmueli K., et al. (2009). Sources of functional magnetic resonance imaging signal fluctuations in the human brain at rest: a 7 T study. Magn. Reson. Imaging 27 1019–1029 10.1016/j.mri.2009.02.004 - DOI - PMC - PubMed
    1. Brett M., Johnsrude I. S., Owen A. M. (2002). The problem of functional localization in the human brain. Nat. Rev. Neurosci. 3 243–249 10.1038/nrn756 - DOI - PubMed
    1. Carlson T. A., Schrater P., He S. (2003). Patterns of activity in the categorical representations of objects. J. Cogn. Neurosci. 15 704–717 10.1162/jocn.2003.15.5.704 - DOI - PubMed

LinkOut - more resources

Feedback