Correlations and anticorrelations in resting-state functional connectivity MRI: A quantitative comparison of preprocessing strategies

Abstract

Resting-state data sets contain coherent fluctuations unrelated to neural processes originating from residual motion artefacts, respiration and cardiac action. Such confounding effects may introduce correlations and cause an overestimation of functional connectivity strengths. In this study we applied several multidimensional linear regression approaches to remove artificial coherencies and examined the impact of preprocessing on sensitivity and specificity of functional connectivity results in simulated data and resting-state data sets from 40 subjects. Furthermore, we aimed at clarifying possible causes of anticorrelations and test the hypothesis that anticorrelations are introduced via certain preprocessing approaches, with particular focus on the effects of regression against the global signal. Our results show that preprocessing in general greatly increased connection specificity, in particular correction for global signal fluctuations almost doubled connection specificity. However, widespread anticorrelated networks were only found when regression against the global signal was applied. Results in simulated data sets compared with result of human data strongly suggest that anticorrelations are indeed introduced by global signal regression and should therefore be interpreted very carefully. In addition, global signal regression may also reduce the sensitivity for detecting true correlations, i.e. increase the number of false negatives. Concluding from our results we suggest that is highly recommended to apply correction against realignment parameters, white matter and ventricular time courses, as well as the global signal to maximize the specificity of positive resting-state correlations.

Introduction

The majority of neuroimaging studies based on blood oxygen level-dependent (BOLD) contrast aim at determining brain areas that are activated by task-specific processing according to a well-defined paradigm. Early work has shown, however, that correlations in low-pass filtered BOLD-weighted data sets acquired under resting-state conditions may also reveal functional networks not related to a specific task (Biswal et al., 1995). While initially demonstrated in cortical motor areas, several resting-state networks have been reported to show strong similarities to well-known functional circuitries including visual areas (Cordes et al., 2000, Kiviniemi et al., 2004, Lowe et al., 1998), auditory cortices (Cordes et al., 2001), language areas (Hampson et al., 2002), basal ganglia (Robinson et al., 2008) and amygdalae (Windischberger et al., 2008). Therefore, it is now widely acknowledged that correlations in spontaneous low-frequency fluctuations relate to known anatomical systems and reflect functional connectivity of the brain (for a review see Fox and Raichle (2007)).

Recently, increasing interest in resting-state functional connectivity MRI has emerged within the neuroscience community related to reports that indicate distinct changes in inter-regional functional connection strength in a variety of psychiatric and neurological disorders: decreased synchrony of low-frequency BOLD fluctuations was observed in bilateral primary motor cortices of multiple sclerosis patients (Lowe et al., 2008, Lowe et al., 2002), in early Alzheimer's disease to the hippocampus (Li et al., 2002, Wang et al., 2006) and between posterior cingulate cortex (PCC) and hippocampus (Greicius et al., 2004), and also within the cortico-limbic network in depression (Anand et al., 2005). In addition, abnormally increased contributions from subgenual cingulate cortex and thalamus in depressive patients (Greicius et al., 2007), as well as altered resting-state functional connectivity of the hippocampus have been reported in schizophrenia patients (Zhou et al., 2008). Resting-state MRI may therefore also serve as an indicator for dysfunctions in brain connectivity, possibly allowing for improved early detection of pathological changes in the brain.

Currently, two major analysis approaches are used for assessing functional connectivity in resting-state data: seed voxel correlation analysis (SCA; Biswal et al., 1995) and independent component analysis (ICA; McKeown et al., 1998). SCA is based on calculating cross-correlation coefficients of the time series in a particular seed region-of-interest (ROI) with all other voxels in the brain, revealing functional connectivity strength with respect to this seed region. SCA thus requires strong a priori assumptions, as the definition of a seed ROI is mandatory. ICA, similar to fuzzy cluster analysis (Moser et al., 1999, Windischberger et al., 2003), represents an exploratory analysis technique decomposing resting-state data sets into different components. As such, standard ICA requires significant operator input in order to distinguish between resting-state networks (RSN), noise components and artefacts. Interestingly, a very promising method for automatic RSN identification based on similarity assessment of individual components in reference to predefined RSN-templates was proposed recently (Greicius et al., 2007).

Besides confirming SCA-derived resting-state networks (RSN), ICA analyses have identified additional RSNs connecting dorsal parietal with lateral prefrontal cortex, precuneus with prefrontal lobe, and inferior temporo-occipito-parietal with inferior prefrontal cortices, respectively (Calhoun et al., 2008, Damoiseaux et al., 2006, De Luca et al., 2006).

Resting-state data sets are scans acquired under baseline conditions, i.e. subjects are instructed to lie still in the scanner, “do nothing” and stay awake. While the definition of this baseline state itself is anything but simple (Stark and Squire, 2001), resting-state fMRI data sets are contaminated by various fluctuations not related to neural activity: residual subject motion, physiological artefacts (caused by respiration, cardiac action), as well as hardware instabilities and magnetic field drifts. Irrespective of whether SCA or ICA is used, functional connectivity analysis basically involves the assessment of temporal coherences in the data set. Signal fluctuations from non-neural processes may introduce coherences and cause an overestimation of functional connectivity strengths. Adequate preprocessing is therefore essential in order to remove such confounding effects prior to the connectivity assessment.

Functional connectivity analysis is performed typically on low-pass filtered data sets with cut-off frequencies of around 0.1 Hz, i.e. below the usual respiration and cardiac action rates of 0.2 and 1 Hz, respectively. Nevertheless, fluctuations from physiological artefacts may still be present, as typical repetition times (TR) of 1 to 2 s prevent critically sampling of cardiac action. Accordingly, cardiac noise is aliased to lower frequency bands and due to the considerable variation in heart rate spread over the whole spectrum (Lund et al., 2006). Therefore, appropriate preprocessing beyond temporal filtering is required, although this fact is ignored in some studies (Li et al., 2002).

One such method allowing for correction against a number of confounds is multidimensional linear regression analysis and has been repeatedly applied to fMRI data sets (Lund et al., 2006). For instance, rigid-body realignment parameters qualify as such regressors, since residual movement artefacts may still be present after realignment and cannot be assigned to a specific frequency range (as would be required for temporal filtering). Effective estimators of non-neural signal fluctuation can also be derived by utilizing MRI-data inherent information, such as signal time courses in white matter (WM). These WM signals are largely independent from BOLD signal changes in cortical and sub-cortical gray matter, but exhibit fluctuations due to scanner instabilities, subject motion and respiration (Windischberger et al., 2002). Physiological artefacts including respiration and cardiac effects are also present in the cerebrospinal fluid (Dagli et al., 1999, Windischberger et al., 2002). Therefore, estimators for physiological noise can be extracted from regions in the ventricular system. It has also been suggested to use the global signal, i.e. the average brain intensity signal, as a nuisance regressor for artefact reduction as it reflects coherent signal fluctuations across the brain (Hampson et al., 2002). While some studies indicate improved fMRI results (Birn et al., 2006), others recommend not to use global scaling in fMRI analysis as it may decrease statistical power (Della-Maggiore et al., 2002). With respect to functional connectivity studies, very recent reports have indicated that regression against the global signal may artificially introduce anticorrelations into fMRI data sets (Murphy et al., 2009, Weissenbacher et al., 2008a, Weissenbacher et al., 2008b). This is of particular importance as some studies employing this regression variant have suggested an intrinsic brain system of anticorrelated networks (Fox et al., 2005).

In this systematic study we aimed to examine the impact of various regression-based correction approaches for resting-state data sets to identify the optimal preprocessing methodology in terms of sensitivity and specificity. In a complementary approach to a very recent study by Murphy et al. (2009) we also wanted to assess possible causes of anticorrelations, with particular focus on the effects caused by regression against the global signal. Therefore, we applied multidimensional linear regression analysis to both simulated data and a group of forty subjects and present a quantitative comparison of the different preprocessing schemes.