Advances and pitfalls in the analysis and interpretation of resting-state FMRI data

Abstract

The last 15 years have witnessed a steady increase in the number of resting-state functional neuroimaging studies. The connectivity patterns of multiple functional, distributed, large-scale networks of brain dynamics have been recognised for their potential as useful tools in the domain of systems and other neurosciences. The application of functional connectivity methods to areas such as cognitive psychology, clinical diagnosis and treatment progression has yielded promising preliminary results, but is yet to be fully realised. This is due, in part, to an array of methodological and interpretative issues that remain to be resolved. We here present a review of the methods most commonly applied in this rapidly advancing field, such as seed-based correlation analysis and independent component analysis, along with examples of their use at the individual subject and group analysis levels and a discussion of practical and theoretical issues arising from this data 'explosion'. We describe the similarities and differences across these varied statistical approaches to processing resting-state functional magnetic resonance imaging signals, and conclude that further technical optimisation and experimental refinement is required in order to fully delineate and characterise the gross complexity of the human neural functional architecture.

Keywords: FMRI; functional connectivity; independent component analysis; networks; resting-state; seed-based correlations.

Figure 1

Eight of the most common and consistent RSNs identified by ICA. (A) RSN located in primary visual cortex; (B) extrastriate visual cortex; (C) auditory and other sensory association cortices; (D) the somatomotor cortex; (E) the ‘default mode’ network (DMN), deactivated during demanding cognitive tasks and involved in episodic memory processes and self-referential mental representations; (F) a network implicated in executive control and salience processing; and (G,H) two right- and left-lateralised fronto-parietal RSNs spatially similar to the bilateral dorsal attention network and implicated in working memory and cognitive attentional processes (for further details, see Beckmann et al., 2005).

Figure 2

Comparison of SCA-derived versions of the DMN using three different seed voxel locations proposed in the literature (A: Fox et al. , in red; B: Singh and Fawcett, in green; C: Greicius et al., , in dark blue). The results of SCA analysis using these seeds are displayed (i) as maximum intensity projections (searching up to 12 voxels below the surface or slice on 3-D renderings of a single subject's high-resolution MRI; RH = right hemisphere, mid = midline, LH = left hemisphere), and (ii) as binarised thresholded Z-statistic images on selected slices in the space of the subject's high resolution MRI (cluster-corrected z = 2.3, p < 0.05). It is clear from the extent of primary (non-overlapping) colours visible (largely red and green), particularly in prefrontal, occipital lobes and subcortical regions, that variations inherent in the seed-selection process can result in a large amount of variability into SCA analysis and subsequent interpretations. (iii) ICA-derived DMN map (Colour bar shows Z-statistic values).

Figure 3

Schematic for temporal concatenation group-ICA.

Figure 4

The vector-space illustration of global mean regression. (A) The characteristic time series for network A can be described as a single point in a high-dimensional vector space. Relative to 0 (the zero time series, black dot) the orthogonal plane (dotted line in this example) separates the vector space into an area of positive correlation (r_A > 0) and a subspace of time series negatively correlated with A. The correlation between A and any other point is defined by the (cosine of) the inner angle: all points within ± 90° are positively correlated with A, whereas all other points are negatively correlated with A; (B) when regressing out the mean of two network-specific time series A and B, the 0 reference point is moved half-way between the two points and the original time series get projected onto the subspace perpendicular to this mean, thereby inducing perfect anti-correlation between A and B as the new characteristic vectors are now aligned at 180°; (C) in the more general case of multiple networks (grey dots) the range of possible differences in pair-wise correlations is again determined by the maximum range of the inner angles α: if α is small, pair-wise correlations differ by only a small amount and delineation of different networks becomes difficult, in this example all pair-wise correlations are positive; (D) the global mean necessarily lies within the convex hull spanned by all the individual characteristic time series. Global time-series regression moves the 0 reference point somewhere into the convex hull, thereby inevitably inducing spurious negative correlations between the characteristic time series associated with different RSNs. Global mean regression does increase the maximum inner angle between pairs of time courses and therefore facilitates delineation of networks from each other; the resulting correlation scores (and signs thereof), however, are no longer interpretable and reference to these should be avoided.

Figure 5

Variability in the strength of inverse coupling between two RSNs (the DMN and a putative executive control network sharing spatial similarity with a combination of regions overlapping with RSN maps from Figures 1F,G,H) associated with individual differences in therapeutic behavioural changes following nicotine pharmacotherapy, compared to placebo. These data are taken from a single subject within a group of smokers tested using resting-state FMRI with repeat measures in a double-blind, placebo-controlled, crossover design (reproduced from Cole et al., under review).