fMRI: Famous vs non-famous faces, SPM

Introduction

The fmri_spm_face.py recreates the classical workflow described in the SPM8 manual using face dataset that can be downloaded from http://www.fil.ion.ucl.ac.uk/spm/data/face_rep/:

python fmri_spm.py

Import necessary modules from nipype.

from __future__ import division
from builtins import range

import nipype.interfaces.io as nio           # Data i/o
import nipype.interfaces.spm as spm          # spm
import nipype.interfaces.matlab as mlab      # how to run matlab
import nipype.interfaces.utility as util     # utility
import nipype.pipeline.engine as pe          # pypeline engine
import nipype.algorithms.modelgen as model   # model specification
import os                                    # system functions

Preliminaries

Set any package specific configuration. The output file format for FSL routines is being set to uncompressed NIFTI and a specific version of matlab is being used. The uncompressed format is required because SPM does not handle compressed NIFTI.

# Set the way matlab should be called
mlab.MatlabCommand.set_default_matlab_cmd("matlab -nodesktop -nosplash")
# If SPM is not in your MATLAB path you should add it here
# mlab.MatlabCommand.set_default_paths('/path/to/your/spm8')

Setting up workflows

In this tutorial we will be setting up a hierarchical workflow for spm analysis. It one is slightly different then the one used in spm_tutorial2.

Setup preprocessing workflow

This is a generic preprocessing workflow that can be used by different analyses

preproc = pe.Workflow(name='preproc')

Use nipype.interfaces.spm.Realign for motion correction and register all images to the mean image.

realign = pe.Node(interface=spm.Realign(), name="realign")

slice_timing = pe.Node(interface=spm.SliceTiming(), name="slice_timing")

Use nipype.interfaces.spm.Coregister to perform a rigid body registration of the functional data to the structural data.

coregister = pe.Node(interface=spm.Coregister(), name="coregister")
coregister.inputs.jobtype = 'estimate'


segment = pe.Node(interface=spm.Segment(), name="segment")
segment.inputs.save_bias_corrected = True

Uncomment the following line for faster execution

# segment.inputs.gaussians_per_class = [1, 1, 1, 4]

Warp functional and structural data to SPM’s T1 template using nipype.interfaces.spm.Normalize. The tutorial data set includes the template image, T1.nii.

normalize_func = pe.Node(interface=spm.Normalize(), name="normalize_func")
normalize_func.inputs.jobtype = "write"

normalize_struc = pe.Node(interface=spm.Normalize(), name="normalize_struc")
normalize_struc.inputs.jobtype = "write"

Smooth the functional data using nipype.interfaces.spm.Smooth.

smooth = pe.Node(interface=spm.Smooth(), name="smooth")

write_voxel_sizes is the input of the normalize interface that is recommended to be set to the voxel sizes of the target volume. There is no need to set it manually since we van infer it from data using the following function:

def get_vox_dims(volume):
    import nibabel as nb
    if isinstance(volume, list):
        volume = volume[0]
    nii = nb.load(volume)
    hdr = nii.header
    voxdims = hdr.get_zooms()
    return [float(voxdims[0]), float(voxdims[1]), float(voxdims[2])]

Here we are connecting all the nodes together. Notice that we add the merge node only if you choose to use 4D. Also get_vox_dims function is passed along the input volume of normalise to set the optimal voxel sizes.

preproc.connect([(realign, coregister, [('mean_image', 'target')]),
                 (coregister, segment, [('coregistered_source', 'data')]),
                 (segment, normalize_func, [('transformation_mat', 'parameter_file')]),
                 (segment, normalize_struc, [('transformation_mat', 'parameter_file'),
                                             ('bias_corrected_image', 'apply_to_files'),
                                             (('bias_corrected_image', get_vox_dims), 'write_voxel_sizes')]),
                 (realign, slice_timing, [('realigned_files', 'in_files')]),
                 (slice_timing, normalize_func, [('timecorrected_files', 'apply_to_files'),
                                                 (('timecorrected_files', get_vox_dims), 'write_voxel_sizes')]),
                 (normalize_func, smooth, [('normalized_files', 'in_files')]),
                 ])

Set up analysis workflow

l1analysis = pe.Workflow(name='analysis')

Generate SPM-specific design information using nipype.interfaces.spm.SpecifyModel.

modelspec = pe.Node(interface=model.SpecifySPMModel(), name="modelspec")

Generate a first level SPM.mat file for analysis nipype.interfaces.spm.Level1Design.

level1design = pe.Node(interface=spm.Level1Design(), name="level1design")

Use nipype.interfaces.spm.EstimateModel to determine the parameters of the model.

level1estimate = pe.Node(interface=spm.EstimateModel(), name="level1estimate")
level1estimate.inputs.estimation_method = {'Classical': 1}

threshold = pe.Node(interface=spm.Threshold(), name="threshold")

Use nipype.interfaces.spm.EstimateContrast to estimate the first level contrasts specified in a few steps above.

contrastestimate = pe.Node(interface=spm.EstimateContrast(), name="contrastestimate")


def pickfirst(l):
    return l[0]

l1analysis.connect([(modelspec, level1design, [('session_info', 'session_info')]),
                    (level1design, level1estimate, [('spm_mat_file', 'spm_mat_file')]),
                    (level1estimate, contrastestimate, [('spm_mat_file', 'spm_mat_file'),
                                                        ('beta_images', 'beta_images'),
                                                        ('residual_image', 'residual_image')]),
                    (contrastestimate, threshold, [('spm_mat_file', 'spm_mat_file'),
                                                   (('spmT_images', pickfirst), 'stat_image')]),
                    ])

Preproc + Analysis pipeline

l1pipeline = pe.Workflow(name='firstlevel')
l1pipeline.connect([(preproc, l1analysis, [('realign.realignment_parameters',
                                            'modelspec.realignment_parameters')])])

Pluging in functional_runs is a bit more complicated, because model spec expects a list of runs. Every run can be a 4D file or a list of 3D files. Therefore for 3D analysis we need a list of lists and to make one we need a helper function.

def makelist(item):
    return [item]
l1pipeline.connect([(preproc, l1analysis, [(('smooth.smoothed_files', makelist),
                                            'modelspec.functional_runs')])])

Data specific components

In this tutorial there is only one subject M03953.

Below we set some variables to inform the datasource about the layout of our data. We specify the location of the data, the subject sub-directories and a dictionary that maps each run to a mnemonic (or field) for the run type (struct or func). These fields become the output fields of the datasource node in the pipeline.

# Specify the location of the data downloaded from http://www.fil.ion.ucl.ac.uk/spm/data/face_rep/face_rep_SPM5.html
data_dir = os.path.abspath('spm_face_data')
# Specify the subject directories
subject_list = ['M03953']
# Map field names to individual subject runs.
info = dict(func=[['RawEPI', 'subject_id', 5, ["_%04d" % i for i in range(6, 357)]]],
            struct=[['Structural', 'subject_id', 7, '']])

infosource = pe.Node(interface=util.IdentityInterface(fields=['subject_id']),
                     name="infosource")

Here we set up iteration over all the subjects. The following line is a particular example of the flexibility of the system. The datasource attribute iterables tells the pipeline engine that it should repeat the analysis on each of the items in the subject_list. In the current example, the entire first level preprocessing and estimation will be repeated for each subject contained in subject_list.

infosource.iterables = ('subject_id', subject_list)

Now we create a nipype.interfaces.io.DataGrabber object and fill in the information from above about the layout of our data. The nipype.pipeline.NodeWrapper module wraps the interface object and provides additional housekeeping and pipeline specific functionality.

datasource = pe.Node(interface=nio.DataGrabber(infields=['subject_id'],
                                               outfields=['func', 'struct']),
                     name='datasource')
datasource.inputs.base_directory = data_dir
datasource.inputs.template = '%s/s%s_%04d%s.img'
datasource.inputs.template_args = info
datasource.inputs.sort_filelist = True

Experimental paradigm specific components

Here we create a structure that provides information about the experimental paradigm. This is used by the nipype.interfaces.spm.SpecifyModel to create the information necessary to generate an SPM design matrix.

from nipype.interfaces.base import Bunch

We’re importing the onset times from a mat file (found on http://www.fil.ion.ucl.ac.uk/spm/data/face_rep/)

from scipy.io.matlab import loadmat
mat = loadmat(os.path.join(data_dir, "sots.mat"), struct_as_record=False)
sot = mat['sot'][0]
itemlag = mat['itemlag'][0]

subjectinfo = [Bunch(conditions=['N1', 'N2', 'F1', 'F2'],
                     onsets=[sot[0], sot[1], sot[2], sot[3]],
                     durations=[[0], [0], [0], [0]],
                     amplitudes=None,
                     tmod=None,
                     pmod=None,
                     regressor_names=None,
                     regressors=None)]

Setup the contrast structure that needs to be evaluated. This is a list of lists. The inner list specifies the contrasts and has the following format - [Name,Stat,[list of condition names],[weights on those conditions]. The condition names must match the names listed in the subjectinfo function described above.

cond1 = ('positive effect of condition', 'T', ['N1*bf(1)', 'N2*bf(1)', 'F1*bf(1)', 'F2*bf(1)'], [1, 1, 1, 1])
cond2 = ('positive effect of condition_dtemo', 'T', ['N1*bf(2)', 'N2*bf(2)', 'F1*bf(2)', 'F2*bf(2)'], [1, 1, 1, 1])
cond3 = ('positive effect of condition_ddisp', 'T', ['N1*bf(3)', 'N2*bf(3)', 'F1*bf(3)', 'F2*bf(3)'], [1, 1, 1, 1])
# non-famous > famous
fam1 = ('positive effect of Fame', 'T', ['N1*bf(1)', 'N2*bf(1)', 'F1*bf(1)', 'F2*bf(1)'], [1, 1, -1, -1])
fam2 = ('positive effect of Fame_dtemp', 'T', ['N1*bf(2)', 'N2*bf(2)', 'F1*bf(2)', 'F2*bf(2)'], [1, 1, -1, -1])
fam3 = ('positive effect of Fame_ddisp', 'T', ['N1*bf(3)', 'N2*bf(3)', 'F1*bf(3)', 'F2*bf(3)'], [1, 1, -1, -1])
# rep1 > rep2
rep1 = ('positive effect of Rep', 'T', ['N1*bf(1)', 'N2*bf(1)', 'F1*bf(1)', 'F2*bf(1)'], [1, -1, 1, -1])
rep2 = ('positive effect of Rep_dtemp', 'T', ['N1*bf(2)', 'N2*bf(2)', 'F1*bf(2)', 'F2*bf(2)'], [1, -1, 1, -1])
rep3 = ('positive effect of Rep_ddisp', 'T', ['N1*bf(3)', 'N2*bf(3)', 'F1*bf(3)', 'F2*bf(3)'], [1, -1, 1, -1])
int1 = ('positive interaction of Fame x Rep', 'T', ['N1*bf(1)', 'N2*bf(1)', 'F1*bf(1)', 'F2*bf(1)'], [-1, -1, -1, 1])
int2 = ('positive interaction of Fame x Rep_dtemp', 'T', ['N1*bf(2)', 'N2*bf(2)', 'F1*bf(2)', 'F2*bf(2)'], [1, -1, -1, 1])
int3 = ('positive interaction of Fame x Rep_ddisp', 'T', ['N1*bf(3)', 'N2*bf(3)', 'F1*bf(3)', 'F2*bf(3)'], [1, -1, -1, 1])

contf1 = ['average effect condition', 'F', [cond1, cond2, cond3]]
contf2 = ['main effect Fam', 'F', [fam1, fam2, fam3]]
contf3 = ['main effect Rep', 'F', [rep1, rep2, rep3]]
contf4 = ['interaction: Fam x Rep', 'F', [int1, int2, int3]]
contrasts = [cond1, cond2, cond3, fam1, fam2, fam3, rep1, rep2, rep3, int1, int2, int3, contf1, contf2, contf3, contf4]

Setting up nodes inputs

num_slices = 24
TR = 2.

slice_timingref = l1pipeline.inputs.preproc.slice_timing
slice_timingref.num_slices = num_slices
slice_timingref.time_repetition = TR
slice_timingref.time_acquisition = TR - TR / float(num_slices)
slice_timingref.slice_order = list(range(num_slices, 0, -1))
slice_timingref.ref_slice = int(num_slices / 2)

l1pipeline.inputs.preproc.smooth.fwhm = [8, 8, 8]

# set up node specific inputs
modelspecref = l1pipeline.inputs.analysis.modelspec
modelspecref.input_units = 'scans'
modelspecref.output_units = 'scans'
modelspecref.time_repetition = TR
modelspecref.high_pass_filter_cutoff = 120

l1designref = l1pipeline.inputs.analysis.level1design
l1designref.timing_units = modelspecref.output_units
l1designref.interscan_interval = modelspecref.time_repetition
l1designref.microtime_resolution = slice_timingref.num_slices
l1designref.microtime_onset = slice_timingref.ref_slice
l1designref.bases = {'hrf': {'derivs': [1, 1]}}

The following lines automatically inform SPM to create a default set of contrats for a factorial design.

# l1designref.factor_info = [dict(name = 'Fame', levels = 2),
#                           dict(name = 'Rep', levels = 2)]

l1pipeline.inputs.analysis.modelspec.subject_info = subjectinfo
l1pipeline.inputs.analysis.contrastestimate.contrasts = contrasts
l1pipeline.inputs.analysis.threshold.contrast_index = 1

Use derivative estimates in the non-parametric model

l1pipeline.inputs.analysis.contrastestimate.use_derivs = True

Setting up parametricvariation of the model

subjectinfo_param = [Bunch(conditions=['N1', 'N2', 'F1', 'F2'],
                           onsets=[sot[0], sot[1], sot[2], sot[3]],
                           durations=[[0], [0], [0], [0]],
                           amplitudes=None,
                           tmod=None,
                           pmod=[None,
                                 Bunch(name=['Lag'],
                                       param=itemlag[1].tolist(),
                                       poly=[2]),
                                 None,
                                 Bunch(name=['Lag'],
                                       param=itemlag[3].tolist(),
                                       poly=[2])],
                           regressor_names=None,
                           regressors=None)]

cont1 = ('Famous_lag1', 'T', ['F2xLag^1'], [1])
cont2 = ('Famous_lag2', 'T', ['F2xLag^2'], [1])
fcont1 = ('Famous Lag', 'F', [cont1, cont2])
paramcontrasts = [cont1, cont2, fcont1]

paramanalysis = l1analysis.clone(name='paramanalysis')

paramanalysis.inputs.level1design.bases = {'hrf': {'derivs': [0, 0]}}
paramanalysis.inputs.modelspec.subject_info = subjectinfo_param
paramanalysis.inputs.contrastestimate.contrasts = paramcontrasts
paramanalysis.inputs.contrastestimate.use_derivs = False

l1pipeline.connect([(preproc, paramanalysis, [('realign.realignment_parameters',
                                               'modelspec.realignment_parameters'),
                                              (('smooth.smoothed_files', makelist),
                                               'modelspec.functional_runs')])])

Setup the pipeline

The nodes created above do not describe the flow of data. They merely describe the parameters used for each function. In this section we setup the connections between the nodes such that appropriate outputs from nodes are piped into appropriate inputs of other nodes.

Use the nipype.pipeline.engine.Pipeline to create a graph-based execution pipeline for first level analysis. The config options tells the pipeline engine to use workdir as the disk location to use when running the processes and keeping their outputs. The use_parameterized_dirs tells the engine to create sub-directories under workdir corresponding to the iterables in the pipeline. Thus for this pipeline there will be subject specific sub-directories.

The nipype.pipeline.engine.Pipeline.connect function creates the links between the processes, i.e., how data should flow in and out of the processing nodes.

level1 = pe.Workflow(name="level1")
level1.base_dir = os.path.abspath('spm_face_tutorial/workingdir')

level1.connect([(infosource, datasource, [('subject_id', 'subject_id')]),
                (datasource, l1pipeline, [('struct', 'preproc.coregister.source'),
                                          ('func', 'preproc.realign.in_files')])
                ])

Setup storage results

Use nipype.interfaces.io.DataSink to store selected outputs from the pipeline in a specific location. This allows the user to selectively choose important output bits from the analysis and keep them.

The first step is to create a datasink node and then to connect outputs from the modules above to storage locations. These take the following form directory_name[.[@]subdir] where parts between [] are optional. For example ‘realign.@mean’ below creates a directory called realign in ‘l1output/subject_id/’ and stores the mean image output from the Realign process in the realign directory. If the @ is left out, then a sub-directory with the name ‘mean’ would be created and the mean image would be copied to that directory.

datasink = pe.Node(interface=nio.DataSink(), name="datasink")
datasink.inputs.base_directory = os.path.abspath('spm_auditory_tutorial/l1output')


def getstripdir(subject_id):
    import os
    return os.path.join(os.path.abspath('spm_auditory_tutorial/workingdir'), '_subject_id_%s' % subject_id)

# store relevant outputs from various stages of the 1st level analysis
level1.connect([(infosource, datasink, [('subject_id', 'container'),
                                        (('subject_id', getstripdir), 'strip_dir')]),
                (l1pipeline, datasink, [('analysis.contrastestimate.con_images', 'contrasts.@con'),
                                        ('analysis.contrastestimate.spmT_images', 'contrasts.@T'),
                                        ('paramanalysis.contrastestimate.con_images', 'paramcontrasts.@con'),
                                        ('paramanalysis.contrastestimate.spmT_images', 'paramcontrasts.@T')]),
                ])

Execute the pipeline

The code discussed above sets up all the necessary data structures with appropriate parameters and the connectivity between the processes, but does not generate any output. To actually run the analysis on the data the nipype.pipeline.engine.Pipeline.Run function needs to be called.

if __name__ == '__main__':
    level1.run()
    level1.write_graph()

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.