NOTE: nii-masker has migrated to nixtract. This repository is no longer maintained. Please use nixtract instead
This is a simple command-line wrapper for nilearn's Masker objects (hence the name), which let you easily extract out region-of-interest (ROI) timeseries from functional MRI data while providing several options for applying additional post-processing (e.g., spatial smoothing, temporal filtering, confound regression, etc). This tool ultimately aims to extend many of nilearn's powerful and convenient masking features to non-Python users (e.g., R and MATLAB users) who wish to analyze fMRI data.
Note: This tool is undergoing rapid early development and is subject to bugs and potentially non-backwards compatible changes.
If you are using this project, please cite nilearn:
Abraham, A., Pedregosa, F., Eickenberg, M., Gervais, P., Mueller, A., Kossaifi, J., … Varoquaux, G. (2014). Machine learning for neuroimaging with scikit-learn. Frontiers in Neuroinformatics, 8, 14.
Link to paper. See also: Nilearn documentation.
niimasker requires the following dependencies:
- numpy
- pandas
- nilearn>=0.5.0
- nibabel
- natsort
- jinja2
First, download this repository to a directory. Then, navigate to the directory, nii-masker/, and run pip install . to install niimasker. To check your installation, run niimasker -h and you should see the help information.
niimasker is run via the command-line and can take the following arguments:
usage: niimasker [-h] [-i input_files [input_files ...]] [-r roi_file]
[-m mask_img] [--labels labels [labels ...]]
[--regressor_files regressor_files [regressor_files ...]]
[--regressors regressors [regressors ...]]
[--regressor_names regressor_names [regressor_names ...]]
[--as_voxels] [--radius radius] [--allow_overlap]
[--standardize] [--t_r t_r] [--high_pass high_pass]
[--low_pass low_pass] [--detrend]
[--smoothing_fwhm smoothing_fwhm]
[--discard_scans discard_scans] [--n_jobs n_jobs]
[-c config]
output_dir
positional arguments:
output_dir The path to the output directory
optional arguments:
-h, --help show this help message and exit
-i input_files [input_files ...], --input_files input_files [input_files ...]
One or more input NIfTI images. Can also be a
single string with a wildcard (*) to specify all
files matching the file pattern. If so, these
files are naturally sorted by file name prior to
extraction.
-r roi_file, --roi_file roi_file
Parameter that defines the region(s) of interest.
This can be 1) a file path to NIfTI image that is
an atlas of multiple regions or a binary mask of
one region, 2) a nilearn query string formatted
as `nilearn:<atlas-name>:<atlas-parameters> (see
online documentation), or 3) a file path to a
.tsv file that has x, y, z columns that contain
roi_file coordinates in MNI space. Refer to
online documentation for more on how these
options map onto the underlying nilearn masker
classes.
-m mask_img, --mask_img mask_img
File path of a NIfTI mask image a to be used when
`roi_file` is a) an multi-region atlas or a b)
list of coordinates. This will restrict
extraction to only voxels within the mask. If
`roi_file` is a single-region binary mask, this
will be ignored.
--labels labels [labels ...]
Labels corresponding to the mask numbers in
`mask`. Can either be a list of strings, or a
.tsv file that contains a `Labels` column. Labels
must be sorted in ascending order to correctly
correspond to the atlas indices. The number of
labels provided must match the number of non-zero
indices in `mask`. If none are provided, numeric
indices are used (default)
--regressor_files regressor_files [regressor_files ...]
One or more tabular files with regressors in each
column. The number of files match the number of
input NIfTI files provided and must be in the
same order. The number of rows in each file must
match the number of timepoints in their
respective input NIfTI files. Can also be a
single string with a wildcard (*) to specify all
files matching the file pattern. If so, these
files are naturally sorted by file name prior to
extraction. Double check to make sure these are
correctly aligned with the input NIfTI files.
--regressors regressors [regressors ...]
Regressor names or strategy to use for confound
regression. Must be a) list of specified column
names in all of the regressor_files, b) a
predefined strategy by load_confounds, or c) a
list compatible with load_confounds flexible
denoising strategy options. See the documentation
https://github.com/SIMEXP/load_confounds. If no
regressor information provided, then all
regressors in regressor_files are used.
--regressor_names regressor_names [regressor_names ...]
Deprecated and will be removed in future
versions. To define specific regressors, please
use --regressors.
--as_voxels Whether to extract out the timeseries of each
voxel instead of the mean timeseries. This is
only available for single ROI binary masks.
Default False.
--radius radius Set the radius of the spheres (in mm) centered on
the coordinates provided in `roi_file`. Only
applicable when a coordinate .tsv file is passed
to `roi_file`; otherwise, this will be ignored.
If not set, the nilearn default of extracting
from a single voxel (the coordinates) will be
used.
--allow_overlap Permit overlapping spheres when coordinates are
provided to `roi_file` and sphere-radius is not
None.
--standardize Whether to standardize (z-score) each timeseries.
Default False
--t_r t_r The TR of the input NIfTI files, specified in
seconds. Must be included if temporal filtering
or realignment derivatives are specified.
--high_pass high_pass
High pass filter cut off in Hertz. If it is not
specified, no filtering is done. (default)
--low_pass low_pass Low pass filter cut off in Hertz.
--detrend Whether to temporally detrend the data.
--smoothing_fwhm smoothing_fwhm
Smoothing kernel FWHM (in mm) if spatial
smoothing is desired.
--discard_scans discard_scans
Discard the first N scans of each functional
NIfTI image.
--n_jobs n_jobs The number of CPUs to use if parallelization is
desired. Default is 1 (serial processing).
-c config, --config config
Configuration .json file as an alternative to
command-line arguments. See online documentation
for what keys to include.
Many of the parameters map directly onto the Masker function arguments in nilearn (see the documentation and user guide for more detail). Additionally, --discard_scans lets you remove the first N scans of your data prior to extraction, --as_voxels lets you get individual voxel timeseries when using a single ROI, and --labels lets you label your ROIs instead of just using the numerical indices. Furthermore, specifying --n_jobs parallelizes the extraction to reduce runtime.
Of course, if you want full nilearn flexibility, you're better off using nilearn and Python directly.
Required parameters
ouput_dir, specified by command-line onlyinput_files, can be specified by the command-line or by a configuration fileroi_file, can be specified by the command-line or by a configuration file
All other arguments are optional.
Say you want to extract the signals from two regions in an image (img1.nii.gz). Region masks are stored in atlas.nii.gz, and have the names "region1" and "region2". With this data, you want to regress out some confounds (e.g., motion realignent parameters, white matter signal, etc. stored in confounds_for_img1.tsv), detrend, high-pass filter, and finally standardize your data. The command to accomplish this is:
niimasker output/ -i img1.nii.gz -r atlas.nii.gz --labels region1 region2 \
--regressor_files confounds_for_img1.tsv --t_r 2 \
--high_pass 0.01 --detrend --standardizeThe averaged BOLD timeseries for region1 and region2 are stored in output/img1_timeseries.tsv, which is a time x region table:
| region1 | region2 |
|---|---|
| .34950 | 1.20285 |
| .82030 | .92329 |
| 1.02455 | 1.51359 |
| .40925 | 1.39835 |
| -.02311 | .22324 |
To keep track over everything, a visual report is generated which shows the extracted timeseries, the region overlays on the averaged functional image, and all the parameters used in the pipeline.
Instead of passing all of the parameters through the command-line, niimasker also provides support for a simple configuration JSON file. The only parameter that needs to be passed into the command-line is the output directory (output_dir). All other parameters can either be set by the configuration file or by the command-line. Note that the configuration file overwrites any of the command-line parameters. An empty configuration file template of all of the parameters is provided in config_template.json, which is shown below:
{
"input_files": [],
"roi_file": "",
"mask_img": null,
"labels": [],
"regressor_files": null,
"regressors": [],
"as_voxels": false,
"sphere_size": null,
"allow_overlap": false,
"standardize": false,
"t_r": null,
"detrend": false,
"high_pass": null,
"low_pass": null,
"smoothing_fwhm": null,
"discard_scans": null,
"n_jobs": 1
}All parameter defaults are shown above. Not all parameters need to be included in the configuration file; only the ones you wish to use. An example use-case that combines both the command-line parameters and configuration file:
niimasker output/ -i img_1.nii.gz img_2.nii.gz -c config.json
Where config.json is:
{
"roi_file": "some_atlas.nii.gz",
"standardize": true,
"regressor_files": [
"confounds1.tsv",
"confounds2.tsv"
],
"regressors": "Params6",
"t_r": 2,
"high_pass": 0.01,
"smoothing_fwhm": 6
}This set up is convenient when your output_dir and input_files vary on a subject-by-subject basis, but your post-processing and atlas might stay constant. Therefore, constants across subjects can be stored in the project's configuration file.
The roi_file parameter (-r or --roi_file) can be a NIfTI image that is either a binary single region mask, or a multi-region atlas wherein each region is uniquely labeled using a numerical value. Or, it can also be a nilearn query string that fetches an atlas via nilearn. Lastly, roi_file can be a .tsv that contains a list of coordinates in the same space of the functional data (e.g., MNI coordinates).
niimasker lets you work with a binary mask such that non-zero values represent a single region of interest mask. Here, you can pass --as_voxels in the CLI or set "as_voxels: true" in your configuration file if you wish to extract the timeseries of every voxel within the region (otherwise the mean timeseries is extracted). This is useful for analyses such as ROI analysis or performing multivariate pattern analyses. A minimal example config.json:
{
"input_files": "img1.nii.gz",
"roi_file": "motor_cortex.nii.gz",
"as_voxels": true,
"standardize": true,
"t_r": 2,
"detrend": true,
"high_pass": 0.01,
}Command:
niimasker output/ -c config.json
If you want to extract out voxelwise data for more than one region, you'll have to provide a new roi_file each time, and thus run niimasker separately for each ROI. Thanks to the configuration file, however, you can just change the command-line call, but keep the configuration file the same. For example, say you have two functional images (each with a TR=2) and you want to extract out two ROIs and perform some temporal filtering. Set the config.json to:
{
"input_files": ["img1.nii.gz", "img2.nii.gz"],
"as_voxels": true,
"standardize": true,
"t_r": 2,
"detrend": true,
"high_pass": 0.01,
"low_pass": null,
"smoothing_fwhm": null,
}Then, you can iteratively call niimasker using a for-loop (python example shown):
import subprocess
rois = ['motor_cortex.nii.gz', 'premotor_cortex.nii.gz']
for mask in rois:
roi_name = mask.split('.')[0] # remove file extension
cmd = ('niimasker timeseries/{} -r {} -c config.json'.format(roi_name, mask))
print(cmd)
subprocess.run(cmd, shell=True)Each iteration changes the ROI and output directory, but the configuration is the same for all ROIs.
niimasker also lets you use an atlas that contains multiple regions, where voxels belonging to each region are labeled with numerical index. This is a typical functional connectivity use-case where you're interested in analyzing the relationships between the timeseries of multiple regions. A typical config.json for two functional images (TR=2), where you want to extract data from a 3-region atlas and perform some additional processing would be:
{
"input_files": ["img1.nii.gz", "img2.nii.gz"],
"roi_file": "some_atlas.nii.gz",
"labels": ["region1", "region2", "region2"],
"regressor_files": [
"confounds1.tsv",
"confounds2.tsv"
],
"regressors": "Params6",
"standardize": true,
"t_r": 2,
"high_pass": 0.01,
"detrend": true,
"discard_scans": 4,
"smoothing_fwhm": 6
}Command:
niimasker output/ -c config.json
Note: When using an atlas, you can also pass a NIfTI image to mask_img (--mask_img or -m) to restrict atlas-extraction to specific voxels. A general example is using a subject specific whole-brain mask in order to exclude non-brain voxels from extraction. Or, you may only want the visual regions of your atlas and you can pass in an occipital lobe mask.
niimasker also gives you the option to specify an atlas that is fetched directly by nilearn (a full list of datasets/atlases that can be fetched by nilearn can be found here). The following atlases are available:
| Atlas Name | Parameter 1 | Parameter 2 | Parameter 3 | nilearn function |
|---|---|---|---|---|
aal |
SPM5, SPM8, SPM12 |
n/a | n/a | fetch_atlas_aal |
basc |
sym, asym |
7, 12, 20, 36, 122, 197, 325, 444 | n/a | fetch_atlas_basc_multiscale_2015 |
destrieux |
lateralized, nonlateralized |
n/a | n/a | fetch_atlas_destrieux_2009 |
schaefer |
100, 200, 300, 400, 500, 600, 700, 800, 900, 1000 | 7, 17 | 1, 2 | fetch_atlas_schaefer_2018 |
talairach |
hemisphere, lobe, gyrus, tissue, ba |
n/a | n/a | fetch_atlas_talairach |
yeo |
thin_7, thick_7, thin_17, thick_17 |
n/a | n/a | fetch_atlas_yeo_2011 |
To fetch an atlas, niimasker takes a query string for roi in the following format: nilearn:<atlas name>:<atlas-parameters>. For instance, the Yeo atlas can be fetch as so: nilearn:yeo:thick_7
Subparameters in the basc and schaefer atlases are separated by hyphen. For instance, nilearn:basc:sym-20 to fetch the Basc symmetrical 20-network atlas, and nilearn:schaefer:400-17-2 to fetch the 400 region, 17-network atlas at 2mm spatial resolution.
Using the above example, we can replace the NIfTI atlas file with the following:
{
"input_files": ["img1.nii.gz", "img2.nii.gz"],
"roi_file": "nilearn:schaefer:400-17-2",
"labels": ["region1", "region2", "region2"],
"regressor_files": [
"confounds1.tsv",
"confounds2.tsv"
],
"regressors": "Params6",
"standardize": true,
"t_r": 2,
"high_pass": 0.01,
"detrend": true,
"discard_scans": 4,
"smoothing_fwhm": 6
}A tab-delimited .tsv file containing coordinates can be passed into roi_file. These coordinates must be in the same space as the functional images (e.g., MNI). The file must contain three columns, and the first row should contain the column names, which must be: 'x', 'y', and 'z'. For instance:
| x | y | z |
|---|---|---|
| 50 | -20 | -42 |
| -38 | -27 | 69 |
niimasker will place a sphere centered around each coordinate, and extract the mean signal of each sphere. The radius of the spheres (in mm) is set using radius (--radius). If no radius is set, then only the signal from the coordinate voxel is extracted. By default, niimasker will raise an error if any of the spheres are overlapping, but you can allow the spheres to overlap by setting allow_overlap: true (or using --allow_overlap). An example configuration file is shown below:
{
"input_files": ["img1.nii.gz", "img2.nii.gz"],
"roi_file": "coordinates.tsv",
"labels": ["region1", "region2"],
"regressor_files": [
"confounds1.tsv",
"confounds2.tsv"
],
"regressors": "Params6",
"radius": 6,
"allow_overlap": true,
"standardize": true,
"t_r": 2,
"high_pass": 0.01,
"detrend": true,
"discard_scans": 4,
"smoothing_fwhm": 6
}Note: A big drawback of placing spheres on coordinates is that the regions are an artificial shape, and will likely include voxels from unwanted sources such as white matter, CSF and non-brain voxels. Passing a gray matter mask to mask_img (-m or --mask_img) will prevent this problem by taking the intersection of the mask and the spheres generated by niimasker/nilearn. Doing so will ensure that each region you are extracting from will contain only gray matter (and therefore many, if not all, spheres will no longer be spheres). It is recommended to use subject-specific masks, such as those generated by Freesurfer. In this case, you will need to run niimasker separately for each subject (see Working with single ROI masks for an example python script that iteratively runs niimasker).
The spheres, or the resulting sphere intersects if mask_img is used, are saved to niimasker_data/spheres_img.nii.gz.
Often times you want to use a set of confound regressors (e.g., head motion parameters) to denoise the timeseries data for connectivity analyses. These confound regressors can be passed into niimasker with the regressor_files parameter, in which each regressor file should have a corresponding functional image in input_files. Both regressor_files and input_files are naturally sorted within niimasker. If these file sets follow the same file naming prefix, they should be in the same order. Always check the corresponding regressor file is used with the input file (see the generated report or parameters.json).
With regressor_files provided, subsets of regressors can be specified using regressors (do not use regressor_names, it is now deprecated). regressors can be column headers of the regressor_files if all headers are present in every file. Or, you can simply pass the name of a predefined confound strategy defined by load_confounds. These currently include:
Params2: Mean white matter and CSF signals, with high-pass filter.Params6: Basic motion parameters with high pass filter.Params9: Basic motion parameters, WM/CSF signals, global signal and high pass filter.Params24: Full motion parameters (derivatives, squares and squared derivatives), with high pass filter.Params36: Motion parameters, WM/CSF signals, global signal, high pass filter. All noise components are fully expanded (derivatives, squares and squared derivatives).AnatCompCor: Motion parameters (fully expanded), high pass filter, and acompcor.TempCompCor: High pass filter, and tcompcor.
Note that regressor files must have the appropriate column names for these strategies to work. If your data is from fmriprep (see below), then the above strategies will automatically work.
niimasker is ideally meant for BIDS-formatted data (although currently not requiring it), and is intended to seamlessly work with fmriprep. To extract data from an entire fmriprep dataset, set your input_files and regressor_files as the following in your config.json:
{
"input_files": "fmriprep/sub*/ses*/func/*preproc_bold.nii.gz",
"regressor_files": "fmriprep/sub*/ses*/func/*confounds_regressors.tsv",
}(Note: Remove /ses*/ if you have single-session data)
Thanks to the BIDS structure of the data, you can provide wildcard patterns for input_files and regressor_files, and niimasker will automatically align regressor files with their respective functional image. From here, you can add in the remaining the parameters you want. An equivalent configuration to the previous example is as follows:
{
"input_files": "fmriprep/sub*/ses*/func/*preproc_bold.nii.gz",
"roi_file": "some_atlas.nii.gz",
"labels": ["region1", "region2", "region2"],
"regressor_files": "fmriprep/sub*/ses*/func/*confounds_regressors.tsv",
"regressors": "Params6",
"standardize": true,
"t_r": 2,
"high_pass": 0.01,
"detrend": true,
"discard_scans": 4,
"smoothing_fwhm": 6
}