Elsevier

NeuroImage

Volume 22, Issue 1, May 2004, Pages 360-366
NeuroImage

A method for removal of global effects from fMRI time series

https://doi.org/10.1016/j.neuroimage.2003.12.042Get rights and content

Abstract

We present a technique for removing global effects from functional magnetic resonance imaging (fMRI) images, using a voxel-level linear model of the global signal (LMGS). The procedure does not assume low-frequency global effects and is based on the assumption that the global signal (the time course of the average intensity per volume) is replicated in the same pattern throughout the brain, although not necessarily at the same magnitude. A second assumption is that all effects that match the global signal are of no interest and can be removed. The method involves modeling the time course of each voxel to the global signal and removing any such global component from the voxel's time course. A challenge that elicits a large change in the global blood oxygenation level-dependent (BOLD) signal, inspired hypercapnia (5% CO2/95% O2), was administered to 14 subjects during a 144-s, 24-scan fMRI procedure; baseline series were also collected. The method was applied to these data and compared to intensity normalization and low-frequency spline detrending. A large global BOLD signal increase emerged to the hypercapnic challenge. Intensity normalization failed to remove global components due to regional variability. Both LMGS and spline detrending effectively removed low-frequency components, but unlike spline detrending (which is designed to remove only low frequency trends), the LMGS removed higher-frequency global fluctuations throughout the challenge and baseline series. LMGS removes all effects correlated with the global signal, and may be especially useful for fMRI data that include large global effects and for generating detrended images to use with subsequent volume-of-interest (VOI) analyses.

Introduction

Functional magnetic resonance imaging (fMRI) is used to detect regional changes in the blood oxygen level-dependent (BOLD) signal corresponding to neuronal responses. These regional signal changes often occur in the presence of global signal changes that are of no interest to the study, and these global effects need to be removed or partitioned in subsequent analyses. A distinction is made between global effects and the global signal; the latter is the average intensity of each brain volume across time. Low-frequency global effects may arise due to scanner drift (Smith et al., 1999) or physiologic fluctuations Biswal et al., 1996, Kiviniemi et al., 2000. Intensity normalization by proportionally scaling each image volume to a constant value is a commonly used method to remove such changes, but this procedure assumes a constant global effect across the brain, which often is not the case (Kastrup et al., 1999). Other techniques effectively adjust for low frequency changes on a voxel-by-voxel basis Kruggel et al., 1999, Tanabe et al., 2002 and thereby account for regional variability. An alternative procedure initially used with PET data is to introduce the global signal as a confounding covariate in the analysis (Friston et al., 1990). These methods are based on the assumption that global changes in the signal are independent of the effects being studied.

Some tasks performed during fMRI studies, such as extreme loaded breathing challenges (Macey et al., 2003), cold pressor evaluation (Harper et al., 2003), and pain-related studies (Tuor et al., 2002), may cause global effects in the BOLD signal due to blood pressure or perfusion alterations. Such task-related global effects usually include a high-frequency component at the onset of the challenge. Intensity normalization would remove such a high-frequency component, but does not account for regional variation; the low-frequency detrending techniques that do account for the local variation in global effects would not remove the higher frequency global components. Alternative methods estimate a “true” global signal that is independent of the experimental paradigm Andersson, 1997, Andersson et al., 2001, Desjardins et al., 2001. The estimated global signal may be used to perform a modified intensity normalization, but this approach rests on the assumption that the estimated global signal is constant across all brain regions. Alternatively, the estimated global signal may be used as a covariate in the analysis Aguirre et al., 1998, Andersson et al., 2001.

For fMRI studies that include large global BOLD effects correlated with experimental paradigms and with more complex responses than standard on–off paradigms, we have experienced difficulties with existing approaches. First, including the global signal as a covariate does not always satisfactorily separate global from experimental effects. We have found that the experimental effect in a linear model may still be contaminated with global effects. Second, negative signals, likely artifactual, are frequently produced. In some cases, these two problems may be resolved by estimating an adjusted global signal Andersson et al., 2001, Desjardins et al., 2001. Third, in contrast to the previous two methods that require a specific model of expected response, we do not wish to restrict the analysis to one model. For studies expected to elicit a sequence of neural responses, such as respiratory or blood pressure challenges, there may be several experimental effects of interest, that is, several models of neural responses. A relatively complex method estimates a global signal independently of the experimental paradigm (Andersson et al., 2001), but this method has not been applied to fMRI data. Finally, volume-of-interest (VOI) analysis, where structures are defined for individual subjects rather than in a template, requires images that are detrended independently of the linear model framework.

We therefore present a technique for removing all global effects from fMRI images, which does not assume that global components are of low frequency in nature, is not based on any particular experimental paradigm, and which allows for regional variation in global effects. The technique is based on a voxel-level linear model of the global signal (LMGS). For the LMGS technique, global effects are assumed to follow the global signal and are assumed to occur in the same pattern throughout the brain, although not necessarily at the same magnitude. A second assumption is that all effects that match the global signal are of no interest and can be removed. The LMGS method removes global effects by modeling the time course of each voxel to the global signal and removing any such global component from the time course of that voxel. The technique was evaluated using a data set of 14 subjects performing a hypercapnic challenge, a task that elicits large global effects in the BOLD signal. The LMGS procedure is compared with intensity normalization and spline detrending.

Section snippets

Subjects and scanning

To induce large global BOLD effects, an inspired hypercapnic gas challenge (5% CO2/95% O2) was administered to 14 healthy children (age: mean, 10.9 years; standard deviation, 2.2 years; range 8–15 years; six female; one left-handed). All studies were performed on a 1.5 Tesla GE scanner using a standard fMRI gradient echo echo-planar imaging (EPI) protocol. Series of 25 EPI volumes [repetition time (TR) = 6000 ms; time to echo (TE) = 60 ms; flip angle = 90°; field of view (FOV) = 30 × 30 cm; no

Regionally variable global increase to hypercapnia

The hypercapnic challenge produced a large increase in global signal which was similar across subjects (Fig. 1). This elevation was distributed throughout the brain [Figs. 2(i–ii)A]. However, the magnitude of signal increase varied regionally. The baseline series had a low-frequency component, which was especially apparent during the first 30 s and is characteristic of scanner drift (Smith et al., 1999).

The global signal of the intensity-normalized whole-brain images was constant. Nevertheless,

Conclusions

Both low- and high-frequency components, and both large-amplitude global effects, such as those induced by hypercapnia, and small-amplitude effects, such as scanner drift, are removed by the LMGS method. This method is based on two assumptions: (1) that the global effect at any voxel follows the same pattern, but not necessarily magnitude, as the global signal; and (2) that any component of the time course of a voxel that follows the pattern of the global signal is of no interest. Therefore,

Acknowledgements

The authors thank Dr. Mary Woo, Ms. Rebecca Harper, and Ms. Amy Kim for their assistance with the data collection. This research was supported by HD-22695.

Cited by (348)

View all citing articles on Scopus
View full text