confound2
¶
confound2
models artifactual signals present in a 4D time series image. The confound model
created by this module can be used to mitigate effects of subject motion and other artifactual
processes by residualizing the 4D image with respect to the confound model. (The
regression/residualization procedure is managed separately in the regress
module.) Several
types of artifact can be modeled: physiological sources, including white matter and CSF signals;
global signal; realignment parameters; and signals derived from principal component analysis (PCA,
CompCor). Derivatives and squares can also be added to the confound model, as can signal during
prior time points. It is a rewrite of the original confound
module to use outputs from
FMRIPREP
.
Model construction order¶
Currently, the confound model is assembled in the following order:
- Add realignment parameters and mean time series from GM, WM, CSF, and global
- Add temporal derivatives of any time series in the model
- Add powers of any time series in the model (e.g., quadratic terms)
- Add component-based time series (CompCor)
- Add custom time series
So, for instance, including the second power will also include not only the squared time series, but also the squares of derivatives.
In the future, a control sequence will probably be implemented to support greater flexibility in confound models.
Module configuration¶
confound2_rps
¶
Realignment parameters
Early models that attempted to correct for the introduction of spurious variance by the movement of subjects in the scanner did so by regressing out the 6 parameters (3 translational, 3 rotational) used to realign each volume in the time series to a reference volume. Later work has demonstrated that a model consisting of realignment parameters alone is ineffective at removing motion artifact from functional MR time series.:
# Use realignment parameters
confound2_rps[cxt]=1
# No realignment parameters
confound2_rps[cxt]=0
confound2_rms
¶
Relative RMS displacement.
The relative root-mean-square displacement is estimated by FSL’s MCFLIRT. This is equivalent to the Jenkinson formulation of framewise displacement and is approximately double the Power formulation of framewise displacement. Using the relative RMS displacement as a confound time series is not recommended; this is an uncommon denoising strategy and is not likely to be effective.:
# Use RMS displacement
confound2_rms[cxt]=1
# No RMS displacement
confound2_rms[cxt]=0
confound2_wm
, and confound2_csf
¶
Tissue-based nuisance regressors, including aCompCor.
Tissue-based nuisance regressors are capable of reducing the influence of subject movement (as well
as physiological artefacts) on the data. Mean white matter and cerebrospinal fluid signal are most
often used to this end (e.g., Windischberger et al., 2002; Satterthwaite et al., 2012), but
principal component analysis can also be used to extract signals of no interest from anatomical
compartments (Behzadi et al., 2007: aCompCor). The number of aCompCor components removed can either
be specified as a fixed number, or by the percent variance explained (usually this is 50% as in
Muschelli et al., 2014.) This approach requires a known segmentation of the
anatomical image into tissue classes. If you provided an output directory from the ANTsCT routine
or the anatomical stream, then a segmentation will automatically be available as the derivative
segmentation
.:
# Do not use any white matter signals
confound2_wm[cxt]=0
# Use the mean white matter signal
confound2_wm[cxt]=1
Interpreting¶
1
indicates that the mean time series over all voxels inside the tissue boundaries should be used in the confound model.
Use of tissue-based nuisance regressors requires a known segmentation of the anatomical image into
tissue classes. If you provided an output directory from the ANTsCT routine or the anatomical
stream, then a segmentation will automatically be available as the derivative segmentation
. In
some segmentations, such as the one output by ANTs Cortical Thickness, each tissue class is
assigned a different intensity value in the segmentation volume. For instance, 1 might correspond
to CSF, 2 to cortical grey matter, 3 to white matter, etc. If your segmentation is strictly a
binary-valued white matter mask, then enter ALL
. To enter a range of values, use the colon
(:
) delimiter; to enter multiple values, use the comma (,
) delimiter.:
# Use a custom segmentation for WM
confound_wm_path[cxt]=/path/to/segmentation.nii.gz
# Use the segmentation from ANTsCT or the anatomical stream for WM
confound_wm_path=${segmentation[sub]}
# Use the mean CSF signal. Use the pipeline segmentation for CSF. 1=CSF in the provided CSF segmentation path.
confound_csf[cxt]=1
confound_csf_path[cxt]=${segmentation[sub]}
confound_csf_val[cxt]=1
In order to ensure that the signal extracted from the tissue or region of interest is not mixed with signal from adjacent voxels associated with a different tissue class (partial volume effects), it is possible to erode its mask by removing fringe voxels. An optimal degree of erosion will result in a mask comprising ‘deep’ voxels of the tissue, while excessive erosion may result in a mask whose extent is poorly representative of the tissue. For functional connectivity analysis, more aggressive erosion of WM and CSF masks is recommended to reduce collinearity of WM and CSF signal with global and GM signals. Erosion to a target range of 5 to 10 percent is recommended in this case.:
# Erode CSF mask to the deepest 10 percent
confound_csf_ero[cxt]=10
# Erode WM mask to the deepest 5 percent
confound_wm_ero[cxt]=5
The value of confound_<tissue>_ero
specifies the level of erosion that is to be applied to
tissue masks. Allowable values range from 0 to 100 and reflect the minimum percentage of tissue
remaining after erosion cycles have been applied. For instance, a value of 30 indicates that the
tissue mask should be eroded to 30 percent its original size; that is, the mask will comprise only
the deepest 30 percent of voxels with the tissue classification. (Depth is computed using
ImageMath
from ANTs, and the erosion is implemented in the utility erodespare
.)
For advanced users: The confound
module offers the option of including up to three tissue- or
RoI-based regressors. While nominally these are the mean GM, WM, and CSF timeseries, it is possible
to include signals from any three RoIs for which a binary mask is available by assigning the
appropriate value to the <tissue>_path
variable.
confound2_gsr
¶
Global signal regression.
Removal of the mean signal across the entire brain is one of the simplest and most effective means of attenuating the influence of artefactual sources such as subject motion. While earlier studies suggested that global signal regression might be harmful, for instance by introducing artefactual anticorrelations (Murphy et al., 2009) or group differences (Saad et al., 2012), an emerging consensus (e.g., Power et al., 2014; Chai et al., 2012) suggests instead that it is uniquely effective in removing widespread forms of artefact (due to both motion and physiological processes such as respiration).:
# Enable GSR (recommended for functional connectivity analysis)
confound2_gsr[cxt]=1
# Disable GSR
confound2_gsr[cxt]=0
confound2_past
¶
Expansion: previous time points.
Including forward-shifted realignment and nuisance timeseries in the nuisance model (Friston et
al., 1996) provides a means of factoring in the subject’s history of motion and for the lingering
effects of motion, which may persist for upwards of 10 seconds following motion itself.
confound_past
must be a nonnegative integer.:
# Include no previous time points
confound2_past[cxt]=0
# Include previous time point
confound2_past[cxt]=1
# Include previous 2 time points
confound2_past[cxt]=2
Note: Do not include both previous time points (confound2_past
) and temporal derivatives
(confound2_dx
) in the same model. Together with the original time series, they form a collinear
triple, which will result in an overspecified model. That is to say, for a time series \(T\), its
temporal derivative \(D\), and previous/shifted time series \(P\),
\(D + P = T\)
confound2_dx
¶
Expansion: temporal derivatives.
Temporal derivatives of motion parameters encode the relative displacement of the brain from one
volume of a timeseries to the next; they are used in major confound models (e.g., Satterthwaite et
al., 2012). confound2_dx
must be a nonnegative integer.:
# Include no temporal derivatives
confound2_dx[cxt]=0
# Include first temporal derivative
confound2_dx[cxt]=1
# Include first and second temporal derivatives
confound2_past[cxt]=2
Note: Do not include both previous time points (confound2_past
) and temporal derivatives
(confound2_dx
) in the same model. Together with the original time series, they form a collinear
triple, which will result in an overspecified model. That is to say, for a time series \(T\), its
temporal derivative \(D\), and previous/shifted time series \(P\),
\(D + P = T\)
confound2_sq
¶
Expansion: powers (quadratic, cubic, quartic, etc.).
In addition to the first power of each confound, you may elect to include higher powers to account for potential noise that is proportional to squares or higher powers of motion parameters and nuisance regressors.:
# First power only
confound2_sq[cxt]=1
# First power and quadratic expansion
confound2_sq[cxt]=2
# First power, quadratic and cubic expansions
confound2_sq[cxt]=3
confound2_custom
¶
Custom regressors.
In addition to regressors generated from the image data, custom regressors can be added to the nuisance model. For instance, these might include respiratory traces convolved with an appropriate response function or estimates of task-driven activity. Custom regressors should be formatted as a matrix with regressor time series in columns and time points/frames in rows.:
# No custom regressors
confound2_custom[cxt]=
# Include a custom regressor file
confound2_custom[cxt]=/path/to/custom/file.1D
# Include custom regressors in multiple files
confound2_custom[cxt]=/path/tocustom/file_1.1D,/path/to/custom/file_2.1D
confound_rerun
¶
Ordinarily, each module will detect whether a particular analysis has run to completion before beginning it. If re-running is disabled, then the module will immediately skip to the next stage of analysis. Otherwise, any completed analyses will be repeated.If you change the run parameters, you should rerun any modules downstream of the change.:
# Skip processing steps if the pipeline detects the expected output
confound_rerun[cxt]=0
# Repeat all processing steps
confound_rerun[cxt]=1
confound2_cleanup
¶
Modules often produce numerous intermediate temporary files and images during the course of an analysis. In many cases, these temporary files are undesirable and unnecessarily consume disk space. If cleanup is enabled, any files stamped as temporary will be deleted when a module successfully runs to completion. If a module fails to detect the output that it expects, then temporary files will be retained to facilitate error diagnosis.:
# Remove temporary files
confound2_cleanup[cxt]=1
# Retain temporary files
confound2_cleanup[cxt]=0
Example configuration: 36-parameters model¶
The 36-parameter confound model includes 6 realignment parameters, mean WM and CSF time series, and
global signal regression (9 parameters). Additionally, the 36-parameter model includes temporal
derivatives of these 9 time series (+9) and squares of the original 9 parameters and of their
temporal derivatives (+18) for a total of 36 parameters. As an illustrative example for
confound2
module configuration, the variable settings for configuring a 36-parameter model are
shown here. The example configuration uses a standard 6-class segmentation, such as that output by
the ANTs Cortical Thickness pipeline when provided appropriate priors.:
confound2_rps[2]=1
confound2_rms[2]=0
confound2_wm[2]=1
confound2_csf[2]=1
confound2_gsr[2]=1
confound2_acompcor[2]=0
confound2_tcompcor[2]=0
confound2_aroma[2]=0
confound2_past[2]=0
confound2_dx[2]=1
confound2_sq[2]=2
confound2_custom[2]=
confound2_censor[2]=0
confound2_censor_contig[2]=0
confound2_framewise[2]=fds:0.167,dv:2
confound2_rerun[2]=0
confound2_cleanup[2]=1