Hands-on 1: How to create a fMRI preprocessing workflow¶
The purpose of this section is that you set-up a complete fMRI analysis workflow yourself. So that in the end you are able to perform the analysis from A-Z, i.e. from preprocessing to group analysis. This section will cover the preprocessing part, and the section Hands-on 2: Analysis will handle the analysis part.
We will use this opportunity to show you some nice additional interfaces/nodes that might not be relevant to your usual analysis. But it's always nice to know that they exist. And hopefully, this will encourage you to investigate all other interfaces that Nipype can bring to the tip of your finger.
Preparation¶
Before we can start with anything we first need to download the data. For this hands-on, we will only use the right-handed subjects 2-4 and 7-9. This can be done very quickly with the following datalad command.
Note: This might take a while, as datalad needs to download ~200MB of data
%%bash
datalad get -J 4 /data/ds000114/sub-0[234789]/ses-test/anat/sub-0[234789]_ses-test_T1w.nii.gz \
/data/ds000114/sub-0[234789]/ses-test/func/*fingerfootlips*
Preprocessing Workflow Structure¶
So let's get our hands dirty. First things first, it's always good to know which interfaces you want to use in your workflow and in which order you want to execute them. For the preprocessing workflow, I recommend that we use the following nodes:
1. Gunzip (Nipype)
2. Drop Dummy Scans (FSL)
3. Slice Time Correction (SPM)
4. Motion Correction (SPM)
5. Artifact Detection
6. Segmentation (SPM)
7. Coregistration (FSL)
8. Smoothing (FSL)
9. Apply Binary Mask (FSL)
10. Remove Linear Trends (Nipype)
Note: This workflow might be overkill concerning data manipulation, but it hopefully serves as a good Nipype exercise.
Imports¶
It's always best to have all relevant module imports at the beginning of your script. So let's import what we most certainly need.
# Get the Node and Workflow object
from nipype import Node, Workflow
# Specify which SPM to use
from nipype.interfaces.matlab import MatlabCommand
MatlabCommand.set_default_paths('/opt/spm12-r7219/spm12_mcr/spm12')
Note: Ideally you would also put the imports of all the interfaces that you use here at the top. But as we will develop the workflow step by step, we can also import the relevant modules as we go.
Create Nodes and Workflow connections¶
Let's create all the nodes that we need! Make sure to specify all relevant inputs and keep in mind which ones you later on need to connect in your pipeline.
Workflow¶
We recommend to create the workflow and establish all its connections at a later place in your script. This helps to have everything nicely together. But for this hands-on example it makes sense to establish the connections between the nodes as we go.
And for this, we first need to create a workflow:
# Create the workflow here
# Hint: use 'base_dir' to specify where to store the working directory
preproc = Workflow(name='work_preproc', base_dir='/output/')
Gunzip¶
I've already created the Gunzip node as a template for the other nodes. Also, we've specified an in_file here so that we can directly test the nodes without worrying about the Input/Output data stream to the workflow. This will be taken care of in a later section.
from nipype.algorithms.misc import Gunzip
# Specify example input file
func_file = '/data/ds000114/sub-07/ses-test/func/sub-07_ses-test_task-fingerfootlips_bold.nii.gz'
# Initiate Gunzip node
gunzip_func = Node(Gunzip(in_file=func_file), name='gunzip_func')
Drop Dummy Scans¶
The functional images of this dataset were recorded with 4 dummy scans at the beginning (see the corresponding publication). But those dummy scans were not yet taken out from the functional images.
To better illustrate this, let's plot the time course of a random voxel of the just defined func_file:
%matplotlib inline
import pylab as plt
import nibabel as nb
plt.plot(nb.load(func_file).get_fdata()[32, 32, 15, :]);
In the figure above, we see that at the very beginning there are extreme values, which hint to the fact that steady state wasn't reached yet. Therefore, we want to exclude the dummy scans from the original data. This can be achieved with FSL's ExtractROI.
from nipype.interfaces.fsl import ExtractROI
extract = Node(ExtractROI(t_min=4, t_size=-1, output_type='NIFTI'),
name="extract")
This ExtractROI node can now be connected to the gunzip_func node from above. To do this, we use the following command:
preproc.connect([(gunzip_func, extract, [('out_file', 'in_file')])])
Slice Time Correction¶
Now to the next step. Let's us SPM's SliceTiming to correct for slice wise acquisition of the volumes. As a reminder, the tutorial dataset was recorded...
- with a time repetition (TR) of 2.5 seconds
- with 30 slices per volume
- in an interleaved fashion, i.e. slice order is [1, 3, 5, 7, ..., 2, 4, 6, ..., 30]
- with a time acquisition (TA) of 2.4167 seconds, i.e.
TR-(TR/num_slices)
from nipype.interfaces.spm import SliceTiming
slice_order = list(range(1, 31, 2)) + list(range(2, 31, 2))
print(slice_order)
# Initiate SliceTiming node here
slicetime = Node(SliceTiming(num_slices=30,
ref_slice=15,
slice_order=slice_order,
time_repetition=2.5,
time_acquisition=2.5-(2.5/30)),
name='slicetime')
Now the next step is to connect the SliceTiming node to the rest of the workflow, i.e. the ExtractROI node.
# Connect SliceTiming node to the other nodes here
preproc.connect([(extract, slicetime, [('roi_file', 'in_files')])])
Motion Correction¶
To correct for motion in the scanner, we will be using FSL's MCFLIRT.
from nipype.interfaces.fsl import MCFLIRT
# Initiate MCFLIRT node here
mcflirt = Node(MCFLIRT(mean_vol=True,
save_plots=True),
name="mcflirt")
Connect the MCFLIRT node to the rest of the workflow.
# Connect MCFLIRT node to the other nodes here
preproc.connect([(slicetime, mcflirt, [('timecorrected_files', 'in_file')])])
Artifact Detection¶
We will use the really cool and useful ArtifactDetection tool from Nipype to detect motion and intensity outliers in the functional images. The interface is initiated as follows:
from nipype.algorithms.rapidart import ArtifactDetect
art = Node(ArtifactDetect(norm_threshold=2,
zintensity_threshold=2,
mask_type='spm_global',
parameter_source='FSL',
use_differences=[True, False],
plot_type='svg'),
name="art")
The parameters above mean the following:
norm_threshold- Threshold to use to detect motion-related outliers when composite motion is being usedzintensity_threshold- Intensity Z-threshold use to detection images that deviate from the meanmask_type- Type of mask that should be used to mask the functional data. spm_global uses an spm_global like calculation to determine the brain maskparameter_source- Source of movement parametersuse_differences- If you want to use differences between successive motion (first element) and intensity parameter (second element) estimates in order to determine outliers
And this is how you connect this node to the rest of the workflow:
preproc.connect([(mcflirt, art, [('out_file', 'realigned_files'),
('par_file', 'realignment_parameters')])
])
Segmentation of anatomical image¶
Now let's work on the anatomical image. In particular, let's use SPM's NewSegment to create probability maps for the gray matter, white matter tissue and CSF.
from nipype.interfaces.spm import NewSegment
# Use the following tissue specification to get a GM and WM probability map
tpm_img ='/opt/spm12-r7219/spm12_mcr/spm12/tpm/TPM.nii'
tissue1 = ((tpm_img, 1), 1, (True,False), (False, False))
tissue2 = ((tpm_img, 2), 1, (True,False), (False, False))
tissue3 = ((tpm_img, 3), 2, (True,False), (False, False))
tissue4 = ((tpm_img, 4), 3, (False,False), (False, False))
tissue5 = ((tpm_img, 5), 4, (False,False), (False, False))
tissue6 = ((tpm_img, 6), 2, (False,False), (False, False))
tissues = [tissue1, tissue2, tissue3, tissue4, tissue5, tissue6]
# Initiate NewSegment node here
segment = Node(NewSegment(tissues=tissues), name='segment')
We will again be using a Gunzip node to unzip the anatomical image that we then want to use as input to the segmentation node. We again also need to specify the anatomical image that we want to use in this case. As before, this will later also be handled directly by the Input/Output stream.
# Specify example input file
anat_file = '/data/ds000114/sub-07/ses-test/anat/sub-07_ses-test_T1w.nii.gz'
# Initiate Gunzip node
gunzip_anat = Node(Gunzip(in_file=anat_file), name='gunzip_anat')
Now we can connect the NewSegment node to the rest of the workflow.
# Connect NewSegment node to the other nodes here
preproc.connect([(gunzip_anat, segment, [('out_file', 'channel_files')])])
Compute Coregistration Matrix¶
As a next step, we will make sure that the functional images are coregistered to the anatomical image. For this, we will use FSL's FLIRT function. As we just created a white matter probability map, we can use this together with the a Boundary-Based Registration (BBR) cost function do optimize the image coregistration. As some helpful notes...
- use a degree of freedom of 6
- specify the cost function as
bbr - use the
schedule='/usr/share/fsl/5.0/etc/flirtsch/bbr.sch'
from nipype.interfaces.fsl import FLIRT
# Initiate FLIRT node here
coreg = Node(FLIRT(dof=6,
cost='bbr',
schedule='/usr/share/fsl/5.0/etc/flirtsch/bbr.sch',
output_type='NIFTI'),
name="coreg")
# Connect FLIRT node to the other nodes here
preproc.connect([(gunzip_anat, coreg, [('out_file', 'reference')]),
(mcflirt, coreg, [('mean_img', 'in_file')])
])
As mentioned above, the bbr routine can use the subject-specific white matter probability map to guide the coregistration. But for this, we need to create a binary mask out of the WM probability map. This can easily be done by FSL's Threshold interface.
from nipype.interfaces.fsl import Threshold
# Threshold - Threshold WM probability image
threshold_WM = Node(Threshold(thresh=0.5,
args='-bin',
output_type='NIFTI'),
name="threshold_WM")
Now, to select the WM probability map that the NewSegment node created, we need some helper function. Because the output field partial_volume_files form the segmentation node, will give us a list of files, i.e. [[GM_prob], [WM_prob], [], [], [], []]. Therefore, using the following function, we can select only the last element of this list.
# Select WM segmentation file from segmentation output
def get_wm(files):
return files[1][0]
# Connecting the segmentation node with the threshold node
preproc.connect([(segment, threshold_WM, [(('native_class_images', get_wm),
'in_file')])])
Now we can just connect this Threshold node to the coregistration node from above.
# Connect Threshold node to coregistration node above here
preproc.connect([(threshold_WM, coreg, [('out_file', 'wm_seg')])])
Apply Coregistration Matrix to functional image¶
Now that we know the coregistration matrix to correctly overlay the functional mean image on the subject-specific anatomy, we need to apply to coregistration to the whole time series. This can be achieved with FSL's FLIRT as follows:
# Specify the isometric voxel resolution you want after coregistration
desired_voxel_iso = 4
# Apply coregistration warp to functional images
applywarp = Node(FLIRT(interp='spline',
apply_isoxfm=desired_voxel_iso,
output_type='NIFTI'),
name="applywarp")
Important: As you can see above, we also specified a variable desired_voxel_iso. This is very important at this stage, otherwise FLIRT will transform your functional images to a resolution of the anatomical image, which will dramatically increase the file size (e.g. to 1-10GB per file). If you don't want to change the voxel resolution, use the additional parameter no_resample=True. Important, for this to work, you still need to define apply_isoxfm=desired_voxel_iso.
# Connecting the ApplyWarp node to all the other nodes
preproc.connect([(mcflirt, applywarp, [('out_file', 'in_file')]),
(coreg, applywarp, [('out_matrix_file', 'in_matrix_file')]),
(gunzip_anat, applywarp, [('out_file', 'reference')])
])
Smoothing¶
Next step is image smoothing. The most simple way to do this is to use FSL's or SPM's Smooth function. But for learning purposes, let's use FSL's SUSAN workflow as it is implemented in Nipype. Note that this time, we are importing a workflow instead of an interface.
from nipype.workflows.fmri.fsl.preprocess import create_susan_smooth
If you type create_susan_smooth? you can see how to specify the input variables to the susan workflow. In particular, they are...
fwhm: set this value to 4 (or whichever value you want)mask_file: will be created in a later stepin_file: will be handled while connection to other nodes in the preproc workflow
# Initiate SUSAN workflow here
susan = create_susan_smooth(name='susan')
susan.inputs.inputnode.fwhm = 4
# Connect Threshold node to coregistration node above here
preproc.connect([(applywarp, susan, [('out_file', 'inputnode.in_files')])])
Create Binary Mask¶
There are many possible approaches on how you can mask your functional images. One of them is not at all, one is with a simple brain mask and one that only considers certain kind of brain tissue, e.g. gray matter.
For the current example, we want to create a dilated gray matter mask. For this purpose we need to:
- Resample the gray matter probability map to the same resolution as the functional images
- Threshold this resampled probability map at a specific value
- Dilate this mask by some voxels to make the mask less conservative and more inclusive
The first step can be done in many ways (eg. using freesurfer's mri_convert, nibabel) but in our case, we will use FSL's FLIRT. The trick is to use the probability mask, as input file and a reference file.
from nipype.interfaces.fsl import FLIRT
# Initiate resample node
resample = Node(FLIRT(apply_isoxfm=desired_voxel_iso,
output_type='NIFTI'),
name="resample")
The second and third step can luckily be done with just one node. We can take almost the same Threshold node as above. We just need to add another additional argument: -dilF - which applies a maximum filtering of all voxels.
from nipype.interfaces.fsl import Threshold
# Threshold - Threshold GM probability image
mask_GM = Node(Threshold(thresh=0.5,
args='-bin -dilF',
output_type='NIFTI'),
name="mask_GM")
# Select GM segmentation file from segmentation output
def get_gm(files):
return files[0][0]
Now we can connect the resample and the gray matter mask node to the segmentation node and each other.
preproc.connect([(segment, resample, [(('native_class_images', get_gm), 'in_file'),
(('native_class_images', get_gm), 'reference')
]),
(resample, mask_GM, [('out_file', 'in_file')])
])
This should do the trick.
Apply the binary mask¶
Now we can connect this dilated gray matter mask to the susan node, as well as actually applying this to the resulting smoothed images.
# Connect gray matter Mask node to the susan workflow here
preproc.connect([(mask_GM, susan, [('out_file', 'inputnode.mask_file')])])
To apply the mask to the smoothed functional images, we will use FSL's ApplyMask interface.
from nipype.interfaces.fsl import ApplyMask
Important: The susan workflow gives out a list of files, i.e. [smoothed_func.nii] instead of just the filename directly. If we would use a normal Node for ApplyMask this would lead to the following error:
TraitError: The 'in_file' trait of an ApplyMaskInput instance must be an existing file name, but a value of ['/output/work_preproc/susan/smooth/mapflow/_smooth0/asub-07_ses-test_task-fingerfootlips_bold_mcf_flirt_smooth.nii.gz'] <class 'list'> was specified.
To prevent this we will be using a MapNode and specify the in_file as it's iterfield. Like this, the node is capable to handle a list of inputs as it will know that it has to apply itself iteratively to the list of inputs.
from nipype import MapNode
# Initiate ApplyMask node here
mask_func = MapNode(ApplyMask(output_type='NIFTI'),
name="mask_func",
iterfield=["in_file"])
# Connect smoothed susan output file to ApplyMask node here
preproc.connect([(susan, mask_func, [('outputnode.smoothed_files', 'in_file')]),
(mask_GM, mask_func, [('out_file', 'mask_file')])
])
Remove linear trends in functional images¶
Last but not least. Let's use Nipype's TSNR module to remove linear and quadratic trends in the functionally smoothed images. For this, you only have to specify the regress_poly parameter in the node initiation.
from nipype.algorithms.confounds import TSNR
# Initiate TSNR node here
detrend = Node(TSNR(regress_poly=2), name="detrend")
# Connect the detrend node to the other nodes here
preproc.connect([(mask_func, detrend, [('out_file', 'in_file')])])
Datainput with SelectFiles and iterables¶
This is all nice and well. But so far we still had to specify the input values for gunzip_anat and gunzip_func ourselves. How can we scale this up to multiple subjects and/or multiple functional images and make the workflow take the input directly from the BIDS dataset?
For this, we need SelectFiles and iterables! It's rather simple, specify a template and fill-up the placeholder variables.
# Import the SelectFiles
from nipype import SelectFiles
# String template with {}-based strings
templates = {'anat': 'sub-{subject_id}/ses-{ses_id}/anat/'
'sub-{subject_id}_ses-test_T1w.nii.gz',
'func': 'sub-{subject_id}/ses-{ses_id}/func/'
'sub-{subject_id}_ses-{ses_id}_task-{task_id}_bold.nii.gz'}
# Create SelectFiles node
sf = Node(SelectFiles(templates,
base_directory='/data/ds000114',
sort_filelist=True),
name='selectfiles')
sf.inputs.ses_id='test'
sf.inputs.task_id='fingerfootlips'
Now we can specify over which subjects the workflow should iterate. To test the workflow, let's still just look at subject 2.
subject_list = ['07']
sf.iterables = [('subject_id', subject_list)]
# Connect SelectFiles node to the other nodes here
preproc.connect([(sf, gunzip_anat, [('anat', 'in_file')]),
(sf, gunzip_func, [('func', 'in_file')])])
Visualize the workflow¶
Now that we're done. Let's look at the workflow that we just created.
# Create preproc output graph
preproc.write_graph(graph2use='colored', format='png', simple_form=True)
# Visualize the graph
from IPython.display import Image
Image(filename='/output/work_preproc/graph.png', width=750)
Run the Workflow¶
Now we are ready to run the workflow! Be careful about the n_procs parameter if you run a workflow in 'MultiProc' mode. n_procs specifies the number of jobs/cores your computer will use to run the workflow. If this number is too high your computer will try to execute too many things at once and will most likely crash.
Note: If you're using a Docker container and FLIRT fails to run without any good reason, you might need to change memory settings in the Docker preferences (6 GB should be enough for this workflow).
preproc.run('MultiProc', plugin_args={'n_procs': 8})
Inspect output¶
What did we actually do? Let's look at all the data that was created.
!tree /output/work_preproc/ -I '*js|*json|*pklz|_report|*dot|*html|*txt|*.m'
But what did we do specifically? Well, let's investigate.
Motion Correction and Artifact Detection¶
How much did the subject move in the scanner and where there any outliers in the functional images?
%matplotlib inline
# Plot the motion paramters
import numpy as np
import pylab as plt
par = np.loadtxt('/output/work_preproc/_subject_id_07/mcflirt/'
'asub-07_ses-test_task-fingerfootlips_bold_roi_mcf.nii.gz.par')
fig, axes = plt.subplots(2, 1, figsize=(15, 5))
axes[0].set_ylabel('rotation (radians)')
axes[0].plot(par[0:, :3])
axes[1].plot(par[0:, 3:])
axes[1].set_xlabel('time (TR)')
axes[1].set_ylabel('translation (mm)');
The motion parameters seems to look ok. What about the detection of artifacts?
# Showing the artifact detection output
from IPython.display import SVG
SVG(filename='/output/work_preproc/_subject_id_07/art/'
'plot.asub-07_ses-test_task-fingerfootlips_bold_roi_mcf.svg')
Which volumes are problematic?
outliers = np.loadtxt('/output/work_preproc/_subject_id_07/art/'
'art.asub-07_ses-test_task-fingerfootlips_bold_roi_mcf_outliers.txt')
list(outliers.astype('int'))
Masks and Probability maps¶
Let's see what all the masks and probability maps look like. For this, we will use nilearn's plot_anat function.
from nilearn import image as nli
from nilearn.plotting import plot_stat_map
%matplotlib inline
output = '/output/work_preproc/_subject_id_07/'
First, let's look at the tissue probability maps.
anat = output + 'gunzip_anat/sub-07_ses-test_T1w.nii'
plot_stat_map(
output + 'segment/c1sub-07_ses-test_T1w.nii', title='GM prob. map', cmap=plt.cm.magma,
threshold=0.5, bg_img=anat, display_mode='z', cut_coords=range(-35, 15, 10), dim=-1);