3.5. Preprocess#

3.5.1. Description#

The Preprocess step of the pipeline is meant for pre-processing before or after Marker step depending on the use-case. For example, you might want to perform confound removal on BOLD data before feature extraction.

This step is still under development and is an optional one for the pipeline to work.

3.5.2. Confound Removal#

The Confound Removal step is meant to remove confounds from the BOLD data. The confounds are extracted from the BOLD_confounds data (must be provided by the Data Grabber). The confounds are then regressed out from the BOLD data using nilearn.image.clean_img().

Currently, junifer supports only one confound removal class: junifer.preprocess.fMRIPrepConfoundRemover. This class is meant to remove confounds as described before, using the output of fMRIPrep as reference.

Strategy#

This confound remover uses the nilearn API from nilearn.interfaces.fmriprep.load_confounds(). That is, define a strategy to extract the confounds from the BOLD_confounds data. The strategy is defined by choosing the noise components to be used and the confounds to be extracted from each noise components. The noise components currently supported are:

motion
wm_csf
global_signal

The confounds options for each noise component are:

basic: the basic confounds for each noise component. For example, for motion, the basic confounds are the 6 motion parameters (3 translations and 3 rotations). For wm_csf, the basic confounds are the mean signal of the white matter and CSF regions. For global_signal, the basic confound is the mean signal of the whole brain.
power2: the basic confounds plus the square of each basic confound.
derivatives: the basic confounds plus the derivative of each basic confound.
full: the basic confounds, the derivative of each basic confound, the square of each basic confound and the square of each derivative of each basic confound.

The strategy is defined as a dictionary, with the noise components as keys and the confounds as values.

Example in python format:

strategy = {
    "motion": "basic",
    "wm_csf": "full",
    "global_signal": "derivatives"
}

or in YAML format:

strategy:
    motion: basic
    wm_csf: full
    global_signal: derivatives

The default value is to use all the noise components with the full confounds:

strategy = {
    "motion": "full",
    "wm_csf": "full",
    "global_signal": "full"
}

Other parameters#

Additionaly, the junifer.preprocess.fMRIPrepConfoundRemover supports the following parameters:

Parameter	Description	Default
`spike`	Add a spike regressor in the timepoints when the framewise displacement exceeds this threshold.	deactivated
`detrend`	Apply detrending on timeseries, before confound removal.	activated
`standardize`	Scale signals to unit variance.	activated
`low_pass`	Low cutoff frequencies, in Hertz.	deactivated
`high_pass`	High cutoff frequencies, in Hertz.	deactivated
`t_r`	Repetition time, in second (sampling period).	from nifti header
`mask`	If provided, signal is only cleaned from voxels inside the mask. If not, a mask is computed using `nilearn.masking.compute_brain_mask()`.	compute