Visual inspection of independent components: Defining a procedure for artifact removal from fMRI data
Introduction
Structured noise from numerous sources (Biswal et al., 1996, Friston et al., 1996, Chen and Zhu, 1997, Birn et al., 1998, Dagli et al., 1999, Glover et al., 2000, Raj et al., 2001, Windischberger et al., 2002, Beauchamp, 2003) and random (Gaussian) noise compromise the functional signal-to-noise ratio and the sensitivity and specificity of analytical results derived from brain blood-oxygenation-level dependent (BOLD) functional magnetic resonance imaging (fMRI) data (Thomas et al., 2002, Huettel et al., 2004a, Raichle and Snyder, 2007). Some structured noise remains in the data after traditional corrections are applied, such as slice-timing correction, rigid-body motion correction, high-pass temporal filtering, and spatial smoothing (Hu et al., 1995, Grootoonk et al., 2000, Andersson et al., 2001, Raj et al., 2001, Birn et al., 2004).
Independent component analysis (ICA) has been used in denoising procedures to improve the sensitivity and specificity of results derived from fMRI data, beyond those obtained with traditional preprocessing (Stone et al., 2002, Thomas et al., 2002, Kochiyama et al., 2005, McKeown et al., 2005, Zou et al., 2009). ICA produces spatiotemporal components (pairs of time courses and spatial maps) through linear decomposition of fMRI data (McKeown et al., 1998). ICA denoising procedures have involved some method for determining (labeling) which independent components (ICs) represent noise (N-ICs) and which represent neural signals of interest (S-ICs). For some studies, the S-ICs have been the denoised end results of interest (Calhoun et al., 2001a, Moritz et al., 2003, Greicius et al., 2007, Stevens et al., 2007, Sui et al., 2009). For other studies, denoising of the fMRI data has been performed as an extension of data preprocessing by (1) filtering the N-ICs from the preprocessed data using the N-IC time courses as nuisance variables with linear regression (Zou et al., 2009); (2) reconstructing the fMRI data from the S-ICs alone (i.e., matrix multiplication of the S-IC time courses and spatial maps) (Thomas et al., 2002, Kochiyama et al., 2005, Perlbarg et al., 2007, Tohka et al., 2008); or (3) projecting (least-squares regression) the fMRI data into the linear subspace spanned by the S-ICs (McKeown, 2000, McKeown et al., 2005). The success of these denoising methods depends upon the accuracy of labeling the ICs, but potential complications in the labeling process are that some ICs appear to represent a synthesis of artifactual and neurally derived signals (McKeown et al., 1998, Thomas et al., 2002, Birn et al., 2008) and it is not clear in every case how ICs should be labeled.
A number of approaches for labeling ICs have been described. These approaches can be divided according to (1) whether ICA is performed on individual fMRI data runs (McKeown, 2000, Calhoun et al., 2001a, Thomas et al., 2002, Moritz et al., 2003, Kochiyama et al., 2005, McKeown et al., 2005, Greicius et al., 2007, Perlbarg et al., 2007, Tohka et al., 2008) or is performed through a single, group ICA on all fMRI data runs together (Stevens et al., 2007, Sui et al., 2009, Zou et al., 2009); (2) whether the approaches are completely automated (McKeown, 2000, Thomas et al., 2002, Kochiyama et al., 2005, McKeown et al., 2005, Greicius et al., 2007, Perlbarg et al., 2007, Stevens et al., 2007, Tohka et al., 2008, Sui et al., 2009) or are “manual” (McKeown et al., 1998, Calhoun et al., 2001a, Moritz et al., 2003, Zou et al., 2009), requiring some element of visual inspection and human decision making; (3) whether the approaches are completely data driven (Calhoun et al., 2001a, Greicius et al., 2007, Perlbarg et al., 2007, Stevens et al., 2007, Tohka et al., 2008, Sui et al., 2009, Zou et al., 2009) or require task-related temporal or spatial (brain areas affected) information (McKeown, 2000, Thomas et al., 2002, Moritz et al., 2003, Kochiyama et al., 2005, McKeown et al., 2005); and (4) whether the approaches are based on IC temporal (McKeown, 2000, Thomas et al., 2002, Kochiyama et al., 2005, McKeown et al., 2005, Greicius et al., 2007, Perlbarg et al., 2007) or spatial (Calhoun et al., 2001a, Stevens et al., 2007, Sui et al., 2009, Zou et al., 2009) characteristics, or both (McKeown et al., 1998, Moritz et al., 2003, Tohka et al., 2008). Characteristics of IC time courses and their associated Fourier frequency spectrums that have been used to distinguish N-ICs from S-ICs include abrupt, large shifts (time course “spikes”) (McKeown et al., 1998, Tohka et al., 2008); oscillating, “quasi-periodic” pattern (McKeown et al., 1998); similarity to white noise (Thomas et al., 2002, Tohka et al., 2008); similarity to time courses from regions of the brain where neural activity does not occur (ventricles and vasculature) (Perlbarg et al., 2007); similarity to task-related activity (McKeown, 2000, Thomas et al., 2002, Moritz et al., 2003, Kochiyama et al., 2005, McKeown et al., 2005); heteroscedacticity in residuals from regressing IC time courses against a task-related time course (Kochiyama et al., 2005); and relative amount of power at frequencies considered typical for artifacts (Thomas et al., 2002, Greicius et al., 2007). IC spatial characteristics used for labeling include degree of association with cerebrospinal fluid, white matter, grey matter, and/or blood vessels (Stevens et al., 2007, Sui et al., 2009, Zou et al., 2009); extent of component variance in brain periphery (McKeown et al., 1998, Tohka et al., 2008); degree of clustering and degree of scattering of thresholded voxels in IC spatial maps (McKeown et al., 1998, Sui et al., 2009); and correspondence with constellations of brain regions known to perform particular functions (Calhoun et al., 2001a, Moritz et al., 2003).
Three recently reported methods for automated labeling of ICs (Perlbarg et al., 2007, Tohka et al., 2008, Sui et al., 2009) were validated in part by comparing the results of automated labeling with those derived from visual inspection. These investigators considered visual inspection by “experts” sufficiently accurate to be used as “a gold standard to assess the quality of the automatic selection procedure” (Perlbarg et al., 2007). However, no operationalized procedure for labeling ICs with visual inspection was described, and such descriptions are lacking in the literature. Perhaps the best description of visual inspection of ICs is found in McKeown et al. (1998), which gives examples of the appearance of spatial and temporal patterns in components that might suggest the presence of artifacts or random noise; but no detailed procedure or guidelines are provided for how to appropriately label components in every case and no data are provided concerning the reliability or accuracy of visual inspection as a method for identifying artifacts.
Detailed descriptions of procedures for visual inspection, ICA-based denoising (VIID) are needed to facilitate further exploration of the potential advantages and disadvantages of using such procedures for denoising fMRI data. For example, it might be useful to explore whether the high levels of accuracy reported for different automated methods, using visual inspection as the standard for comparison, would be as high if automated labeling results were compared to visual inspection results at a location with slightly different visual inspection traditions or philosophies. Such exploration would require a more detailed methodology description than “visual inspection was performed.” Here, we provide an example of a detailed description of visual inspection of ICs as part of a procedure for denoising fMRI data.
Section snippets
Overview
This procedure derives a set of ICs, each representing a separate portion of the total variance in the fMRI data, using spatial ICA. We assume that the sources of variance for each component are a mixture of neural signals of interest (NSI), neural signals of no interest (NSNI, e.g., activity related to a brain function or region not being studied), structured noise, and random noise. The primary goal of this procedure is to reduce noise in the fMRI data, while preserving as much NSI as
Results
For the resting-state data, the individual-ICA pass resulted in 21–37 components for each subject, for a total of 464 components; 292 (63%) were labeled N-IC. In the second pass, 2 (13%) of 16 components were labeled N-IC. Inter-rater agreement was 96%, with Cohen's κ = 0.91. The total time required by SM for calibration sessions was 165 min; and for the labeling procedure 45 min, corresponding to an average of 9 min per subject and 22 s per IC. Little time was required to move from one IC to the
Discussion
The main purpose of this article is to advocate for detailed descriptions of what is meant by visual inspection of IC spatial maps and/or time courses, in articles that use this term. Such descriptions are needed to facilitate the exploration of visual inspection as part of fMRI data denoising procedures. The level of detail in such descriptions must be sufficient to allow reproduction by independent investigators, meaningful comparison of findings, and scientific discussion of the potential
Acknowledgments
This work was supported by NIMH grants R01 MH065653, P30 MH085943, T32 MH019132 (to George S. Alexopoulos, M.D.), K23 MH067702 (to Christopher F. Murphy, Ph.D.), K23 MH074818 (to Faith M. Gunning, Ph.D.), the Sanchez and TRU Foundations, and Forest Pharmaceuticals, Inc. Dr. Alexopoulos has received research grants by Forest Pharmaceuticals, Inc. and Cephalon and participated in scientific advisory board meetings of Forest Pharmaceuticals. He has given lectures supported by Forest, Bristol
References (64)
- et al.
Modeling geometric deformations in EPI time series
Neuroimage
(2001) - et al.
Experimental designs and processing strategies for fMRI studies involving overt verbal responses
Neuroimage
(2004) - et al.
Separating respiratory-variation-related fluctuations from neuronal-activity-related fluctuations in fMRI
Neuroimage
(2006) - et al.
fMRI activation in a visual-perception task: network of areas detected using the general linear model and independent components analysis
Neuroimage
(2001) - et al.
Group independent component analysis reveals consistent resting-state networks across multiple sessions
Brain Res
(2008) - et al.
Localization of cardiac-induced signal change in fMRI
Neuroimage
(1999) - et al.
Resting-state functional connectivity in major depression: abnormally increased contributions from subgenual cingulate cortex and thalamus
Biol Psychiatry
(2007) - et al.
The effect of artifacts on dependence measurement in fMRI
Magn Reson Imaging
(2006) - et al.
Characterization and correction of interpolation effects in the realignment of fMRI time series
Neuroimage
(2000) - et al.
Improved optimization for the robust and accurate linear registration and motion correction of brain images
Neuroimage
(2002)