Elsevier

NeuroImage

Volume 112, 15 May 2015, Pages 267-277
NeuroImage

ICA-AROMA: A robust ICA-based strategy for removing motion artifacts from fMRI data

https://doi.org/10.1016/j.neuroimage.2015.02.064Get rights and content

Highlights

  • ICA-AROMA: ICA-based strategy for motion artifact removal from fMRI data.

  • ICA-AROMA preserves degrees of freedom, prevents heteroscedasticity.

  • ICA-AROMA is generalizable across datasets, does not require re-training.

  • ICA-AROMA is applicable to resting-state fMRI and task-based fMRI data.

Abstract

Head motion during functional MRI (fMRI) scanning can induce spurious findings and/or harm detection of true effects. Solutions have been proposed, including deleting (‘scrubbing’) or regressing out (‘spike regression’) motion volumes from fMRI time-series. These strategies remove motion-induced signal variations at the cost of destroying the autocorrelation structure of the fMRI time-series and reducing temporal degrees of freedom. ICA-based fMRI denoising strategies overcome these drawbacks but typically require re-training of a classifier, needing manual labeling of derived components (e.g. ICA-FIX; Salimi-Khorshidi et al. (2014)). Here, we propose an ICA-based strategy for Automatic Removal of Motion Artifacts (ICA-AROMA) that uses a small (n = 4), but robust set of theoretically motivated temporal and spatial features. Our strategy does not require classifier re-training, retains the data's autocorrelation structure and largely preserves temporal degrees of freedom. We describe ICA-AROMA, its implementation, and initial validation. ICA-AROMA identified motion components with high accuracy and robustness as illustrated by leave-N-out cross-validation. We additionally validated ICA-AROMA in resting-state (100 participants) and task-based fMRI data (118 participants). Our approach removed (motion-related) spurious noise from both rfMRI and task-based fMRI data to larger extent than regression using 24 motion parameters or spike regression. Furthermore, ICA-AROMA increased sensitivity to group-level activation. Our results show that ICA-AROMA effectively reduces motion-induced signal variations in fMRI data, is applicable across datasets without requiring classifier re-training, and preserves the temporal characteristics of the fMRI data.

Introduction

Head motion during functional Magnetic Resonance Imaging (fMRI) scanning results in misalignment of one volume to the next. This introduces measurement inaccuracies as imaging voxels do not represent identical brain regions over time. Primary effects of participant head motion in fMRI data are corrected by realigning volumes using linear alignment algorithms. However, head motion does not only result in misaligned volumes, but also induces secondary effects related to partial voluming, interpolation effects, magnetic field inhomogeneities, intra-volume motion, and spin-history effects (Friston et al., 1996, Beall and Lowe, 2014), which cannot be corrected for by using volume-realignment. The most common strategy to correct for these secondary effects is to model participant head motion and remove the modeled responses from the fMRI data using additional linear regressors within the framework of the General Linear Model (GLM; Friston et al. (1996)).

Recent findings have spurred renewed debate on the impact of participant head motion on resting state fMRI (rfMRI) experiments particularly. Most prominently, participant head motion during an rfMRI measurement could induce spurious temporal correlation between brain regions, even in light of generally adopted strategies for motion-induced artifact correction (Van Dijk et al., 2012, Satterthwaite et al., 2012, Power et al., 2012). Functional connectivity measures derived from rfMRI could be particularly affected by spurious temporal correlations as they investigate temporal correlations in the absence of a task-related model. Accordingly, in light of increased head motion in younger participants or the possibility of motion as a disease trait (e.g. patients with attention-deficit/hyperactive disorder), discriminating signal from noise when investigating developmental or disease-related neural signatures will be complicated by interactions with head motion (Van Dijk et al., 2012, Satterthwaite et al., 2013b).

In contrast to model-free functional connectivity analyses, typical task-based fMRI analyses investigate BOLD activity related to an experimental model and realignment-based measurements of head motion can be included as covariates. However, in addition to ‘spontaneous movement’, task-based fMRI analyses can be affected by stimulus-related motion, which cannot be separated from stimulus-related signal variations of interest in the regression model. Accordingly, regression-based motion artifact removal strategies potentially remove signal of interest, thereby decreasing sensitivity to functional activation.

The most common strategy for dealing with secondary effects of participant head motion in fMRI data is nuisance regression. Typical nuisance regression models include 6 to 24 motion-related covariates derived from volume-realignment parameters (Friston et al., 1996, Yan et al., 2013a; Satterthwaite et al., 2013a). Six or 12 motion-related covariates were initially considered sufficient, but currently 24 covariates are recommended (Yan et al., 2013a, Satterthwaite et al., 2013a). In addition, and specifically for rfMRI data, recent strategies have proposed excluding volumes associated with high motion from the fMRI time-series (Power et al., 2012, Power et al., 2014, Satterthwaite et al., 2013a). Two (analogous) strategies accomplish this goal by respectively regressing out (‘spike regression’; Lemieux et al., 2007, Satterthwaite et al., 2013a) or deleting individual high-motion volumes from the fMRI time-series (‘scrubbing’; Power et al., (2012)).

The strategies above aim to rigorously implement correction for motion-related artifacts, yet several drawbacks can be identified. First, each strategy relies on the realignment parameters (RPs) obtained from realigning volumes during primary correction for motion in fMRI data. Naturally, these parameters can only be as accurate as the algorithm used for realignment. Second, the regression-based strategies typically only model linear motion-induced signal variation while the underlying dynamics are non-linear (Fair et al., 2012), i.e. secondary effects of head motion are not necessarily captured by the obtained realignment parameters. Third, the use of a large set of nuisance regressors may lead to overfitting of the data and therefore to removal of signal of interest (Yan et al., 2013a, Satterthwaite et al., 2013a). Fourth, by removing specific high-motion volumes, spike regression and scrubbing destroy the autocorrelation structure of the data. This will impact frequency filtering typically employed within fMRI preprocessing (Carp, 2013) and prevent any analysis that is aimed at investigating frequency characteristics (e.g. amplitude low-frequency fluctuations; ALFF) or non-stationarity in functional connectivity (Yan et al., 2013a). Finally, such volume removal or regression results in a high and variable loss of temporal degrees of freedom (tDoF; Yan et al. (2013a)). As an example, in two reported high-motion cohorts, the mean amount of deleted volumes was respectively 26% and 58% (Power et al., 2012). The number of available volumes is typically regarded as the available number of tDoF, making it clear that the associated loss in statistical power induced by spike regression or scrubbing can be substantial and can differ substantially between subsets of subjects after volume removal/regression. Importantly, tDoF determine the estimation accuracy of subject-level statistics. Reduced tDoF for instance affect the error variance within a typical single-subject, first-level regression. Accordingly, although spike regression and scrubbing can reduce the association between motion artifact and measures of interest across a population (Yan et al., 2013b), they can introduce an association between the amount of motion and the accuracy of single-subject statistics. This, in turn, introduces heteroscedasticity at the between-subject level.

To avoid such drawbacks, Yan et al. (2013a) suggested employing of extensive nuisance regression, including 24 RPs, at the single-subject level and including motion covariates in group-level analyses. However, group-level motion covariates can share variance with variables of interest and therefore reduce sensitivity to an effect of interest. As an example, motion might act as a trait (Van Dijk et al., 2012, Couvy-Duchesne et al., 2014, Kong et al., 2014) and have a neural basis (Pujol et al., 2014, Zeng et al., 2014). Moreover, group-level covariates do not correct for inferior estimation of single-subject effect sizes in the presence of high levels of noise at the subject level.

Finally, although controversial, global signal regression and band-pass filtering are frequently considered for confound removal. Global signal regression reduces the impact of motion on functional connectivity metrics (Satterthwaite et al., 2013a, Yan et al., 2013a) but it inevitably removes signal of interest as the global signal is a superposition of both signal and noise components. Additionally, global signal regression introduces anti-correlations and alters connectivity structure (Murphy et al., 2009, Weissenbacher et al., 2009, Yan et al., 2013b). Similarly, temporal band-pass filtering removes signal of interest present at higher frequencies (Niazy et al., 2011, Liao et al., 2013, Kalcher et al., 2014).

The drawbacks of current strategies to remove motion-induced signal variations from fMRI data are at least partly addressed by alternative strategies that aim to identify and remove motion-related artifacts using Independent Component Analysis (ICA). Applied to fMRI data, ICA decomposes the data into a set of spatial independent component maps (ICs), and associated time-courses (McKeown et al., 1998, Beckmann and Smith, 2004, Beckmann et al., 2005). The resulting components represent brain activity, and/or structured noise (e.g. motion-related, physiological or scanner-induced noise). Components representing noise can be regressed out from the data, implying that ICA can be used to remove noise from fMRI data in a data-driven fashion (Thomas et al., 2002, Kelly et al., 2010, Kundu et al., 2012). However, labeling ICA components as noise or signal of interest is a subjective and time-consuming process. Multiple methods have been developed to automatically identify noise components based on temporal and/or spatial features (Thomas et al., 2002, Kochiyama et al., 2005, De Martino et al., 2007, Perlbarg et al., 2007, Tohka et al., 2008, Kundu et al., 2012, Bhaganagarapu et al., 2013, Rummel et al., 2013, Storti et al., 2013, Salimi-Khorshidi et al., 2014). These methods have not been widely adopted due to lack of accuracy or extensive validation over multiple datasets. Some strategies have specific disadvantages such as being only applicable to task-based fMRI data (Kochiyama et al., 2005) or multi-echo fMRI data (Kundu et al., 2012), being limited to physiological noise (Thomas et al., 2002, Perlbarg et al., 2007) or requiring to re-train the classifier for every new dataset acquired using a different MR scanner and/or MRI protocol (Thomas et al., 2002, De Martino et al., 2007, Salimi-Khorshidi et al., 2014).

To address current issues associated with both motion parameter- and ICA-based strategies for motion artifact removal from fMRI data, we propose an alternative ICA-based strategy. In contrast to generic ICA-based denoising strategies (Salimi-Khorshidi et al., 2014), we here focus specifically on the classification and removal of components that specifically relate to head motion. In the light of an increasing number of large-scale multi-site studies, we aim to develop a robust strategy that does not require classifier re-training across datasets. To that end we construct a classifier using a limited and theoretically motivated set of features. These features are defined a priori and correspond to component characteristics which are typically evaluated during manual classification. Note that such an approach substantially differs from an approach in which a complex classifier is trained using an extensive set of features, inherently increasing the probability of a biased classifier towards the training dataset. Specifically, our strategy implements a classifier that employs two temporal and two spatial features. The ICA-based denoising is applicable to rfMRI and task-based fMRI data, largely preserves the autocorrelation structure of the fMRI time-series, and has little impact on the tDoF, thereby avoiding heteroscedasticity in group-level statistics.

In the first section of this manuscript we discuss our ICA-based strategy, its features, classifier construction, and development. The second section comprises an evaluation of the classifier and a validation study in which we applied our strategy to rfMRI and task-based fMRI data. In both datasets we investigated the removal of group-level spurious noise, sensitivity to activation, and loss in tDoF. Throughout all assessments we compared results obtained using ICA-AROMA to preprocessing with extensive nuisance regression (Satterthwaite et al., 2013a, Yan et al., 2013a) and spike regression (Satterthwaite et al., 2013a). Of note, a complete evaluation of ICA-AROMA against alternative strategies for removing motion-related artifacts was beyond the scope of this manuscript. For such an evaluation we refer the reader to a companion manuscript where we compared nine strategies by assessing the achieved quality of motion artifact removal, preservation of signal of interest, and replication across multiple rfMRI datasets (Pruim et al., 2015).

Section snippets

ICA-AROMA

Fig. 1 provides an overview of our ICA-based strategy for motion artifact removal called ‘ICA-AROMA’ or ‘ICA-based Automatic Removal of Motion Artifacts’. Within the typical fMRI participant-level preprocessing stream ICA-AROMA is applied after spatial smoothing but prior to high-pass filtering and further nuisance regression. ICA-AROMA includes three consecutive steps. The first step is a probabilistic ICA on the partly preprocessed single-subject fMRI data. Next, ICA-AROMA exploits a set of

Classifier evaluation

To evaluate the ICA-AROMA classifier we assessed its classification accuracy in the training dataset and the robustness of the classifier by means of a leave-N-out cross-validation. When applying the ICA-AROMA classifier on the full set of components of the training set, including all 30 participants, ICA-AROMA removed a mean of 23.1 out of 34.5 components across participants (see Table 2). Fig. 3 presents the classification results specifically for each manually defined class; RSN, motion and

Discussion

We proposed ICA-AROMA, a novel strategy for removing motion artifacts from fMRI data. ICA-AROMA exploits independent component analysis to identify participant-specific motion-related components in a data-driven fashion. These components are subsequently removed from the data. We showed that ICA-AROMA effectively removed motion artifacts from both rfMRI and task-based fMRI data, while increasing sensitivity to signal of interest. By avoiding the removal of fMRI volumes, ICA-AROMA largely

Conclusion

We have developed a robust and automated strategy to identify and remove motion-related artifact from fMRI data. Our strategy employs independent component analysis, and a simple, four-feature classifier, that does not require re-training in new datasets. ICA-AROMA is applicable to both rfMRI and task-based fMRI data, effectively removing (motion-related) spurious noise while increasing sensitivity to activation. An initial validation showed that ICA-AROMA outperforms extensive nuisance

Acknowledgments

The NeuroIMAGE project was supported by NWO Large Investment Grant 1750102007010, ZonMW Grant 60-60600-97-193, and NWO Brain and Cognition grants 433-09-242 and 056-13-015 to Dr. Buitelaar, the EU FP7 grant TACTICS (grant no. 278948), and grants from Radboud University Medical Center, University Medical Center Groningen and Accare, and VU University Amsterdam. Dr. Mennes has received funding from the European Research Council under the European Union's Seventh Framework Programme (FP7/2007-2013

Conflict of interest

The authors declare that they have no conflicts of interest.

References (57)

  • M. Jenkinson et al.

    Improved optimization for the robust and accurate linear registration and motion correction of brain images

    NeuroImage

    (2002)
  • M. Jenkinson et al.

    FSL

    NeuroImage

    (2012)
  • R.E. Kelly et al.

    Visual inspection of independent components: defining a procedure for artifact removal from fMRI data

    J. Neurosci. Methods

    (2010)
  • T. Kochiyama et al.

    Removing the effects of task-related motion using independent-component analysis

    NeuroImage

    (2005)
  • P. Kundu et al.

    Differentiating BOLD and non-BOLD signals in fMRI time series using multi-echo EPI

    NeuroImage

    (2012)
  • L. Lemieux et al.

    Modelling large motion events in fMRI studies of patients with epilepsy

    Magn. Reson. Imaging

    (2007)
  • X.H. Liao et al.

    Functional brain hubs and their test–retest reliability: a multiband resting-state functional MRI study

    NeuroImage

    (2013)
  • K. Murphy et al.

    The impact of global signal regression on resting state correlations: are anti-correlated networks introduced?

    NeuroImage

    (2009)
  • R.K. Niazy et al.

    Spectral characteristics of resting state networks

    Prog. Brain Res.

    (2011)
  • V. Perlbarg et al.

    CORSICA: correction of structured noise in fMRI by automatic identification of ICA components

    Magn. Reson. Imaging

    (2007)
  • J.D. Power et al.

    Spurious but systematic correlations in functional connectivity MRI networks arise from subject motion

    NeuroImage

    (2012)
  • J.D. Power et al.

    Methods to detect, characterize, and remove motion artifact in resting state fMRI

    NeuroImage

    (2014)
  • J. Pujol et al.

    Does motion-related brain functional connectivity reflect both artifacts and genuine neural activity?

    NeuroImage

    (2014)
  • G. Salimi-Khorshidi et al.

    Automatic denoising of functional MRI data: combining independent component analysis and hierarchical fusion of classifiers

    NeuroImage

    (2014)
  • T.D. Satterthwaite et al.

    Impact of in-scanner head motion on multiple measures of functional connectivity: relevance for studies of neurodevelopment in youth

    NeuroImage

    (2012)
  • T.D. Satterthwaite et al.

    An improved framework for confound regression and filtering for control of motion artifact in the preprocessing of resting-state functional connectivity data

    NeuroImage

    (2013)
  • T.D. Satterthwaite et al.

    Heterogeneous impact of motion on fundamental patterns of developmental changes in functional connectivity during youth

    NeuroImage

    (2013)
  • S.M. Smith et al.

    Advances in functional and structural MR image analysis and implementation as FSL

    NeuroImage

    (2004)
  • Cited by (0)

    View full text