regress

regress executes multiple linear regression to fit any confound time series computed using the confound module to the time series of each voxel in the analyte image. Any variance in the BOLD time series that is explained by the confound model is discarded from the analyte image. The residuals (unexplained variance) of the BOLD time series are retained as the denoised image. regress additionally supports temporal filtering, censoring, and production of smoothed derivative time series.

Despiking

Despiking is a process in which large spikes in the BOLD times series are truncated. Despiking reduces/limits the amplitude or magnitude of the large spikes but preserves those data points with an imputed reduced amplitude. Despiking is encouraged to be done before filtering and regression to minimize the impact of spike. Despiking is very effective because it is a voxelwise operation, and no volume is deleted/removed. For despiking in xcpEngine, DSP is added to regress_process:

regress_process[cxt]=DSP-TMP-REG

Temporal Censoring

Temporal Censoring is a process in which data points with excessive motion outliers are identified/flagged. The censored data points are removed from the data. This is effective for removing spurious sources of connectivity in fMRI data but must be applied verycarefully because the censored volumes are removed and the final BOLD signal. To apply censoring in xcpEngine, confound2_censor[cxt]=1 should be specified in confound2 in the design file. However, the threshold to identify the censored volumes or outliers is obtained from framewise displacement. The framewise displacement (FD) is obtained from FMRIPREP regressors.

Framewise displacement (fd) will be used as the threshold. For instance, the spike regression (Satterthwaite et al. 2013) required threshold of 0.25mm for TR of 3s and scrubbing (Power 2012) required threshold of 0.5 mm. Any volume with FD above the threshold will be flagged. The threshold is configured in confound2 as shown below. Scrubbing/censoring uses FD,RMS and DVARS to established the threshold. The threshold is specified in confound2 as well as:

  • confound2_framewise[cxt]=fds:0.083,dv:2. For spike regression, with TR of of 3, the threshold will be

0.083*3 ~= 0.25mm. with This is standard or common threshold (Ciric et al. 2012, Satterthwaite et al. 2013) and are provided in the design files. The actual threshold value is determined by the TR of the BOLD signal.

For scrubbing, the threshold will be fds:0.167 which implies that the actual threshold will be 0.167*3 =~0.5mm Scrubbing also includes masking out the non-contiguous segments of data between outliers. The number of contiguous volumes required to survive masking is set flexibly by censor_contig[cst]=x, where x is the minimum number of contiguous threshold-surviving time points required for those time points to survive masking. The default found in non-scrubbing design files is zero. This can be configured by setting:

*confound2_censor_contig[cxt]=5

Surface processing

THe xcpEngine also regresses out the confound regressors from the gifti and cifti files if they are available in the fmriprep output. The residualised files are as follows:

  • prefix_residualized.dtseries # residualised cifti files
  • prefix_residualized_hemi-R_bold.func.gii # residualized gifti files

regress_process

Specifies the order for execution of filtering and regression. Bandpass filtering the analyte time series but not nuisance regressors re-introduces noise-related variance at removed frequencies when the time series is residualised with respect to the regressors via linear fit (Hallquist et al., 2014). Thus, effective denoising requires either that confound regression be performed prior to temporal filtering or that both the analyte time series and all confound time series be subjected to the same temporal filter in order to prevent frequency mismatch. Format regress_process as a string of concatenated three-character routine codes separated by hyphens (-).

  • REG-TMP instructs the module to perform confound regression prior to temporal filtering.
  • TMP-REG instructs the module to perform temporal filtering prior to confound regression. If this option is set, then both the analyte time series and all confound time series will be filtered. This option is typically preferable to REG-TMP because it determines the confound fit using only the frequencies of interest.
  • Note that censoring is always performed after both filtering and regression.

regress_tmpf

Temporal filtering parameters.

Bandpass filtering the analyte time series but not nuisance regressors re-introduces noise-related variance at removed frequencies when the time series is residualised with respect to the regressors via linear fit (Hallquist et al., 2014). (The XCP Engine is designed so as to make this involuntary reintroduction of noise impossible.) Instead, the recommended approach is filtering both the time series and the nuisance regressors immediately prior to fitting and residualisation (Hallquist et al., 2014).:

# Gaussian filter
regress_tmpf[cxt]=gaussian

# FFT filter
regress_tmpf[cxt]=fft

# Second-order Butterworth filter
regress_tmpf[cxt]=butterworth
regress_tmpf_order[cxt]=2
regress_tmpf_pass[cxt]=2

# First-order Chebyshev I filter with pass-band ripple 0.5
regress_tmpf[cxt]=chebyshev1
regress_tmpf_order[cxt]=1
regress_tmpf_pass[cxt]=2
regress_tmpf_ripple[cxt]=0.5

# First-order elliptic filter with pass-band ripple 0.5 and stop-band ripple 20
regress_tmpf[cxt]=elliptic
regress_tmpf_order[cxt]=1
regress_tmpf_pass[cxt]=2
regress_tmpf_ripple[cxt]=0.5
regress_tmpf_ripple2[cxt]=20

To note:

  • FFT-based filters, as implemented in AFNI’s 3dBandpass, use a fast Fourier transform to attenuate frequencies. An FFT-based filter may not be suitable for use in designs that incorporate iterative motion censoring, since it will include interpolated frequencies in its calculations.
  • A Gaussian filter, as implemented in FSL, uses a Gaussian-weighted least-squares fit to remove frequencies of no interest from the data. This filter has a very slow frequency roll-off.
  • Chebyshev and elliptic filters more ideally discriminate accepted and attenuated frequencies than do Butterworth filters, but they introduce ripples in either the passband (Chebyshev I), stopband (Chebyshev II), or both (elliptic) that result in some signal distortion.
  • regress*tmpf*order specifies the filter order. (Relevant only for Butterworth, Chebyshev, and elliptic filters.)
  • regress*tmpf*pass specifies whether the filter is forward-only (regress*tmpf*pass[cxt]=1, analogous to filter or lfilter in NumPy or MATLAB) or forward-and-reverse (regress*tmpf*pass[cxt]=2, analogous to filtfilt in NumPy or MATLAB, recommended). (Relevant only for Butterworth, Chebyshev, and elliptic filters.)
  • regress*tmpf*ripple specifies the pass-band ripple, while regress*tmpf*ripple2 specifies the stop-band ripple. (ripple relevant only for Chebyshev I or elliptic filter, ripple2 relevant only for Chebyshev II or elliptic filter.)

regress_hipass and regress_lopass

Temporal filter cutoff frequencies.

Any frequencies below the low-pass cutoff and above the high-pass cutoff will be counted as pass-band frequencies; these will be retained by the filter when it is applied.

Functional connectivity between regions of interest is typically determined on the basis of synchrony in low-frequency fluctuations (Biswal et al., 1995); therefore, removing higher frequencies using a low-pass filter may effectively remove noise from the time series while retaining signal of interest. For a contrasting view, see Boubela et al. (2013). Set regress_lopass to n (Nyquist) to allow all low frequencies to pass.:

# Band-pass filter with pass-band 0.01-0.08 Hz
regress_hipass[cxt]=0.01
regress_lopass[cxt]=0.08

# High-pass-only filter (>0.01 Hz)
regress_hipass[cxt]=0.01
regress_lopass[cxt]=n

# Low-pass-only filter (<0.1 Hz)
regress_hipass[cxt]=0
regress_lopass[cxt]=0.1

High-pass filters can be used to remove very-low-frequency drift from an acquisition; this is a form of scanner noise. The demean/detrend option additionally removes linear and polynomial drift. Set regress_hipass to 0 to allow all high frequencies to pass.

regress_sptf and regress_smo: Spatial smoothing parameters.

Endemic noise, for instance due to physiological signals or scanner activity, can introduce spurious or artefactual results in single voxels. The effects of noise-related artefacts can be mitigated by spatially filtering the data, thus dramatically increasing the signal-to-noise ratio. However, spatial smoothing is not without its costs: it effectively reduces volumetric resolution by blurring signals from adjacent voxels. The spatial smoothing implemented in the regress module (i) keeps the unsmoothed analyte image for downstream use and (ii) creates a derivative image that is smoothed using the specified kernel. This allows either the smoothed or the unsmoothed version of the image to be used in any downstream modules as appropriate.:

# No smoothing
regress_sptf[cxt]=none
regress_smo[cxt]=0

# Gaussian kernel (fslmaths) of FWHM 6 mm
regress_sptf[cxt]=gaussian
regress_smo[cxt]=6

# SUSAN kernel (FSL's SUSAN) of FWHM 4 mm
regress_sptf[cxt]=susan
regress_smo[cxt]=4

# Uniform kernel (AFNI's 3dBlurToFWHM) of FWHM 5 mm
regress_sptf[cxt]=uniform
regress_smo[cxt]=5

regress_sptf specifies the type of spatial filter to apply for smoothing, while regress_smo specifies the full-width at half-maximum (FWHM) of the smoothing kernel in mm.

  • Gaussian smoothing applies the same Gaussian smoothing kernel across the entire volume.
  • SUSAN-based smoothing restricts mixing of signals from disparate tissue classes (Smith and Brady, 1997).
  • Uniform smoothing applies smoothing to all voxels until the smoothness computed at every voxel attains the target value.
  • Uniform smoothing may be used as a compensatory mechanism to reduce the effects of subject motion on the final processed image (Scheinost et al., 2014).

regress_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
regress_rerun[cxt]=0

# Repeat all processing steps
regress_rerun[cxt]=1

regress_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
regress_cleanup[cxt]=1

# Retain temporary files
regress_cleanup[cxt]=0

Expected outputs

The main output of regress module is prefix_residualised.nii.gz for the completion of the module. Other outputs include:

-  prefix_confmat.1D  # filtered regressors
-  prefix_confcor.txt # Pearson correlation between confound regressors

For Censoring and spike regression and if any volume is flagged, the other outputs include:

- prefix_uncensored.nii.gz # the regressed bold image with  flagged volme  interpolated
- prefix_nVolumesCensored.txt # number of volume censored
- prefix_residualised.nii.gz # residualized volume with deleted flagged volume

The optional output is the spatially smoothed residualised BOLD signal. This is specified with regress_sptf[cxt] and regress_smo[cxt] as explained previously. For instance, with:

regress_sptf[cxt]=gaussian
regress_smo[cxt]=6

This tells xcpEngine to smooth the residualised image with gaussian filter and kernel of 6mm. The derived output, which is saved in the regress folder, will be:

prefix_img_sm6.nii.gz

If the freesurfer is included as part of FMRIPREP outputs, the CIFTI files are produced for filtered and regressed bold data (${prefix}_residualised.nii.gz)