Workshop: Dartmouth College 2010

First lets go to the directory with the data we’ll be working on and start the interactive python interpreter (with some nipype specific configuration). Note that nipype does not need to be run through ipython - it is just much nicer to do interactive work in it.

cd $TDPATH
ipython -p nipype

For every neuroimaging procedure supported by nipype there exists a wrapper - a small piece of code managing the underlying software (FSL, SPM, AFNI etc.). We call those interfaces. They are standarised so we can hook them up together. Lets have a look at some of them.

In [1]: import nipype.interfaces.fsl as fsl

In [2]: fsl.BET.help()
Inputs
------

Mandatory:
 in_file: input file to skull strip

Optional:
 args: Additional parameters to the command
 center: center of gravity in voxels
 environ: Environment variables (default={})
 frac: fractional intensity threshold
 functional: apply to 4D fMRI data
  mutually exclusive: functional, reduce_bias
 mask: create binary mask image
 mesh: generate a vtk mesh brain surface
 no_output: Don't generate segmented output
 out_file: name of output skull stripped image
 outline: create surface outline image
 output_type: FSL output type
 radius: head radius
 reduce_bias: bias field and neck cleanup
  mutually exclusive: functional, reduce_bias
 skull: create skull image
 threshold: apply thresholding to segmented brain image and mask
 vertical_gradient: vertical gradient in fractional intensity threshold (-1, 1)

Outputs
-------
mask_file: path/name of binary brain mask (if generated)
meshfile: path/name of vtk mesh file (if generated)
out_file: path/name of skullstripped file
outline_file: path/name of outline file (if generated)

In [3]: import nipype.interfaces.freesurfer as fs

In [4]: fs.Smooth.help()
Inputs
------

Mandatory:
 in_file: source volume
 num_iters: number of iterations instead of fwhm
  mutually exclusive: surface_fwhm
 reg_file: registers volume to surface anatomical
 surface_fwhm: surface FWHM in mm
  mutually exclusive: num_iters
  requires: reg_file

Optional:
 args: Additional parameters to the command
 environ: Environment variables (default={})
 proj_frac: project frac of thickness a long surface normal
  mutually exclusive: proj_frac_avg
 proj_frac_avg: average a long normal min max delta
  mutually exclusive: proj_frac
 smoothed_file: output volume
 subjects_dir: subjects directory
 vol_fwhm: volumesmoothing outside of surface

Outputs
-------
args: Additional parameters to the command
environ: Environment variables
smoothed_file: smoothed input volume
subjects_dir: subjects directory

You can read about all of the interfaces implemented in nipype at our online documentation at http://nipy.sourceforge.net/nipype/documentation.html#documentation . Check it out now.

Using interfaces

Having interfaces allows us to use third party software (like FSL BET) as function. Look how simple it is.

from __future__ import print_function
from builtins import str

import nipype.interfaces.fsl as fsl
result = fsl.BET(in_file='data/s1/struct.nii').run()
print(result)

Running a single program is not much of a breakthrough. Lets run motion correction followed by smoothing (isotropic - in other words not using SUSAN). Notice that in the first line we are setting the output data type for all FSL interfaces.

fsl.FSLCommand.set_default_output_type('NIFTI_GZ')
result1 = fsl.MCFLIRT(in_file='data/s1/f3.nii').run()
result2 = fsl.Smooth(in_file='f3_mcf.nii.gz', fwhm=6).run()

Simple workflow

In the previous example we knew that fsl.MCFLIRT will produce a file called f3_mcf.nii.gz and we have hard coded this as an input to fsl.Smooth. This is quite limited, but luckily nipype supports joining interfaces in pipelines. This way output of one interface will be used as an input of another without having to hard code anything. Before connecting Interfaces we need to put them into (separate) Nodes and give them unique names. This way every interface will process data in a separate folder.

import nipype.pipeline.engine as pe
import os

motion_correct = pe.Node(interface=fsl.MCFLIRT(in_file=os.path.abspath('data/s1/f3.nii')),
                         name="motion_correct")
smooth = pe.Node(interface=fsl.Smooth(fwhm=6), name="smooth")

motion_correct_and_smooth = pe.Workflow(name="motion_correct_and_smooth")
motion_correct_and_smooth.base_dir = os.path.abspath('.')  # define where will be the root folder for the workflow
motion_correct_and_smooth.connect([
    (motion_correct, smooth, [('out_file', 'in_file')])
])
# we are connecting 'out_file' output of motion_correct to 'in_file' input of smooth
motion_correct_and_smooth.run()

Another workflow

Another example of a simple workflow (calculate the mean of fMRI signal and subtract it). This time we’ll be assigning inputs after defining the workflow.

calc_mean = pe.Node(interface=fsl.ImageMaths(), name="calc_mean")
calc_mean.inputs.op_string = "-Tmean"
subtract = pe.Node(interface=fsl.ImageMaths(), name="subtract")
subtract.inputs.op_string = "-sub"

demean = pe.Workflow(name="demean")
demean.base_dir = os.path.abspath('.')
demean.connect([
    (calc_mean, subtract, [('out_file', 'in_file2')])
])

demean.inputs.calc_mean.in_file = os.path.abspath('data/s1/f3.nii')
demean.inputs.subtract.in_file = os.path.abspath('data/s1/f3.nii')
demean.run()

Reusing workflows

The beauty of the workflows is that they are reusable. We can just import a workflow made by someone else and feed it with our data.

from fmri_fsl import preproc
preproc.base_dir = os.path.abspath('.')
preproc.inputs.inputspec.func = os.path.abspath('data/s1/f3.nii')
preproc.inputs.inputspec.struct = os.path.abspath('data/s1/struct.nii')
preproc.run()

... and we can run it again and it won’t actually rerun anything because none of the parameters have changed.

preproc.run()

... and we can change a parameter and run it again. Only the dependent nodes are rerun and that too only if the input state has changed.

preproc.inputs.meanfuncmask.frac = 0.5
preproc.run()

Visualizing workflows 1

So what did we run in this precanned workflow

preproc.write_graph()

Datasink

Datasink is a special interface for copying and arranging results.

import nipype.interfaces.io as nio

preproc.inputs.inputspec.func = os.path.abspath('data/s1/f3.nii')
preproc.inputs.inputspec.struct = os.path.abspath('data/s1/struct.nii')
datasink = pe.Node(interface=nio.DataSink(), name='sinker')
preprocess = pe.Workflow(name='preprocout')
preprocess.base_dir = os.path.abspath('.')
preprocess.connect([
    (preproc, datasink, [('meanfunc2.out_file', 'meanfunc'),
                         ('maskfunc3.out_file', 'funcruns')])
])
preprocess.run()

Datagrabber

Datagrabber is (surprise, surprise) an interface for collecting files from hard drive. It is very flexible and supports almost any file organisation of your data you can imagine.

datasource1 = nio.DataGrabber()
datasource1.inputs.template = 'data/s1/f3.nii'
datasource1.inputs.sort_filelist = True
results = datasource1.run()
print(results.outputs)

datasource2 = nio.DataGrabber()
datasource2.inputs.template = 'data/s*/f*.nii'
datasource2.inputs.sort_filelist = True
results = datasource2.run()
print(results.outputs)

datasource3 = nio.DataGrabber(infields=['run'])
datasource3.inputs.template = 'data/s1/f%d.nii'
datasource3.inputs.sort_filelist = True
datasource3.inputs.run = [3, 7]
results = datasource3.run()
print(results.outputs)

datasource4 = nio.DataGrabber(infields=['subject_id', 'run'])
datasource4.inputs.template = 'data/%s/f%d.nii'
datasource4.inputs.sort_filelist = True
datasource4.inputs.run = [3, 7]
datasource4.inputs.subject_id = ['s1', 's3']
results = datasource4.run()
print(results.outputs)

Iterables

Iterables is a special field of the Node class that enables to iterate all workfloes/nodes connected to it over some parameters. Here we’ll use it to iterate over two subjects.

import nipype.interfaces.utility as util
infosource = pe.Node(interface=util.IdentityInterface(fields=['subject_id']),
                     name="infosource")
infosource.iterables = ('subject_id', ['s1', 's3'])

datasource = pe.Node(nio.DataGrabber(infields=['subject_id'], outfields=['func', 'struct']), name="datasource")
datasource.inputs.template = '%s/%s.nii'
datasource.inputs.base_directory = os.path.abspath('data')
datasource.inputs.template_args = dict(func=[['subject_id', 'f3']], struct=[['subject_id', 'struct']])
datasource.inputs.sort_filelist = True

my_workflow = pe.Workflow(name="my_workflow")
my_workflow.base_dir = os.path.abspath('.')

my_workflow.connect([(infosource, datasource, [('subject_id', 'subject_id')]),
                     (datasource, preproc, [('func', 'inputspec.func'),
                                            ('struct', 'inputspec.struct')])])
my_workflow.run()

and we can change a node attribute and run it again

smoothnode = my_workflow.get_node('preproc.smooth')
assert(str(smoothnode) == 'preproc.smooth')
smoothnode.iterables = ('fwhm', [5., 10.])
my_workflow.run()

Visualizing workflows 2

In the case of nested workflows, we might want to look at expanded forms of the workflow.

Example source code

You can download the full source code of this example. This same script is also included in the Nipype source distribution under the examples directory.