Research articleImages-based suppression of unwanted global signals in resting-state functional connectivity studies
Introduction
The assessment of functional connectivity between brain areas relies on the identification of regions that show covariance, or correlation, in the evolution of BOLD fMRI signal, usually acquired in a steady state, at rest or during continuous stimulation. Coherent temporal fluctuations of the BOLD signal in distinct brain regions, observed especially at relatively low temporal frequencies (typical corresponding period of 10–100 s), have been explained as a consequence of simultaneous neuronal engagement and finally as an indication that the involved regions are connected in a functional network [1]. Resting-state networks have been identified, initially in the motor cortices [2] then in several other brain areas [3]. The concept of resting networks partially overlap with the default mode network, introduced to explain both the deactivations observed especially in the anterior and posterior cingulate during a wide range of demanding tasks, and the relatively low increase of brain energy consumption during activation if compared to the baseline consumption [4], [5]. The later report that large-scale correlations persist during the execution of continuous cognitive or somatosensory tasks, and during light sedation, led to the hypothesis that the brain is intrinsically organized in collaborating networks, irrespectively of the conditions of rest or task execution [6], [7], [8].
Studies aimed at identifying regions functionally connected capitalize on fluctuations of the BOLD signal that are usually rejected as correlated noise in the conventional fMRI analysis. Instrumental noise can play a role in this context, given that BOLD-like low-frequency fluctuations have been reported in water phantoms and in cadavers [9], [10]; however, instrumental noise was found to be of less impact than physiological noise at 3 T [11] but not necessarily at higher fields [12], [13]. Particularly harmful is the correlated signal related to cardiac pulsation and respiration, and to subject motion [14], [15], [16]. Cardiac rhythm is undersampled by the typical fMRI scans, that sample the signal at a rate of about 0.5 s−1, corresponding to a Nyquist limit of 0.25 s−1 in frequency (or 4 s in period) for whole brain (WB) coverage. Cardiac pulsatility noise is found especially near the brain edges or in proximity of vessels or sinuses [16], [17]; however, gray matter (GM) is generally prone to cardiac noise, probably due to the high vascularization [18], [19]. The noise spectrum depends on the degree of aliasing; however, it usually shows components within the band of interest for functional connectivity magnetic resonance imaging (fcMRI) analyses [17], [19], [20]. Besides the bulk head movement, that is accounted for by the movement realignment, at least for sufficiently slow and regular inspirations, respiration effects manifest as distortions induced by the magnetic field modulation induced by chest movement and volume variations and as modulation of blood oxygenation [20], [21]. They have a generally widespread spatial distribution [16], [17], but the cingulate cortex has been found especially prone to blood CO2 modulations [16], [21]. While the main respiratory rhythm, around 0.2 Hz, is usually critically sampled and above the usual cutoff frequency for fcMRI analyses, volume and rhythm modulations show slower components, in the range of some dozen of seconds, thus acting as confounds in fcMRI analyses [16]. Finally, motion-related noise, as is well known, is a feature of high-contrast boundaries, especially at the brain edges.
In order to reduce the impact of these confounding factors, several approaches have been proposed. The most successful are based on the use of simultaneously recorded physiological parameters from which a model of the physiological noise is derived. The model is then fitted and subtracted from the fMRI series [16], [17] or is equivalently introduced as nuisance variables of a general linear model (GLM) fitting [20]. Usually, respiratory and cardiac rhythms and movement realignment parameters, as well as terms accounting for the modulation of those rhythms, and for their possible undersampling at the low sampling rate typical of fMRI series, are included in the model. However, these approaches have the drawback of needing the simultaneous recording of physiological parameters because only the movement parameters can be retrospectively estimated from the images themselves.
A possible alternative approach is based on the extraction and subsequent subtraction of global signals of noninterest (GSNI) from the images. This is based usually on independent component analysis (ICA) decomposition or on regression/correlation with regions of noninterest. ICA is attractive because it is in principle able to identify global signals independent from the brain activity without a priori assumptions; however, it usually relies on the posterior visual identification of the component to be discarded [22], [23], although approaches for automatic classification of components have been proposed, based on specific prior knowledge of the correlated noise characteristics [24], [25]. While theoretically suited for steady-state analyses, the effects of ICA filtering on fcMRI analyses were only recently and only partially characterized [26]. Compared with ICA, regression of GSNI exploits the extraction of average signal from voxels that are a priori assumed to carry mainly noise and minimally signal of interest. Usually, the signal from white matter (WM) and blood vessels, as well as the global signal from the WB or from each slice, are assumed to well represent the GSNI [7], [27], [28]. This method was successfully used in fcMRI studies [7], [29]. However, it suffers two main and related drawbacks: first, the chosen signals can carry a substantial part of the signals of interest; second, they can be only partially representative of GSNI. On the whole, the procedure of subtraction/regression can suffer both low specificity and low sensitivity.
The main goals of this study were (1) to investigate the spatial and temporal characteristics of GSNI, as extracted with a linear regression approach, and (2) to assess the changes in fcMRI maps related to the subtraction of those GSNI. The experiment is based on resting-state fMRI scans in order to avoid possible interactions with the execution of a task. Our results suggest that subtraction based on regression of GSNI is a sufficiently accurate way to manage the physiological noise in fcMRI studies at 3 T.
Section snippets
Subjects and data acquisition
Fifteen healthy, right-handed subjects (three females, mean 27, S.D.7 years) participated in this study after giving informed consent, according to the national laws and to the local ethics committee guidelines. MR data were acquired on a 3-T Siemens Magnetom Allegra head scanner, equipped with fast (400 mT m−1 ms−1) and strong (40 mT/m) gradients, using a circularly polarized, standard head coil. Subjects were immobilized with foam pillows. A T1-weighted sagittal image was acquired for
Results
A substantial amount of variance at voxel level (in the order of 20%) was explained by slow (putatively instrumental) drifting in all the data sets. Several voxels showed a significant correlation with the GSNI regressors; the intersubject average ratio of voxels showing significant GSNI components is reported in Table 1. The spatial distribution of voxels correlated with WM and CSF signal in a representative subject, as well as the relevant binary tissue masks, is shown in Fig. 1. Voxels
Discussion
Conventional fMRI detects essentially the brain response to a specific challenge, and with this approach, an incredibly large amount of knowledge about brain function, either physiological or pathological, have been reached. However, it is intrinsically insensitive to modulations of the baseline. On the other hand, fcMRI is characterized by the ability of mapping the baseline “activity,” via the analysis of low-frequency BOLD fluctuations. This feature is attractive for the study of the healthy
References (35)
- et al.
Empirical analyses of BOLD fMRI statistics. I. spatially unsmoothed data collected under null-hypothesis conditions
Neuroimage
(1997) - et al.
Investigation of low frequency drift in fMRI signal
Neuroimage
(1999) - et al.
Comparison of physiological noise at 1.5 T, 3 T and 7 T and optimization of fMRI acquisition parameters
Neuroimage
(2005) - et al.
Sources of fMRI signal fluctuations in the human brain at rest: a 7T study
Magn Reson Imaging
(2009) - et al.
Separating respiratory-variation-related fluctuations from neuronal-activity-related fluctuations in fMRI
Neuroimage
(2006) - et al.
Localization of cardiac-induced signal change in fmri
Neuroimage
(1999) - et al.
Low-frequency fluctuations in the cardiac rate as a source of variance in the resting-state fMRI BOLD signal
Neuroimage
(2007) - et al.
Non-white noise in fMRI: does modelling have an impact?
Neuroimage
(2006) - et al.
Resting fluctuations in arterial carbon dioxide induce significant low frequency variations in BOLD signal
Neuroimage
(2004) - et al.
Noise reduction in BOLD-based fMRI using component analysis
Neuroimage
(2002)