Source code for cortex.mni

"""
Functions for finding MNI transforms for individual subjects and transforming
functional data and surfaces to and from MNI space.
"""

import os
import nibabel
import tempfile
import subprocess
import numpy as np

from . import options
from . import db

import shlex

fslprefix = options.config.get("basic", "fsl_prefix")
fsldir = os.getenv("FSLDIR")
if fsldir is None:
    import warnings
    warnings.warn("Can't find FSLDIR environment variable, assuming default FSL location..")
    fsldir = "/usr/share/fsl/5.0"

default_template = os.path.join(fsldir, "data", "standard", "MNI152_T1_1mm_brain.nii.gz")

def _save_fsl_xfm(filename, xfm):
    np.savetxt(filename, xfm, "%0.10f")

def _load_fsl_xfm(filename):
    return np.loadtxt(filename)

[docs] def compute_mni_transform(subject, xfm, template=default_template): """ Compute transform from the space specified by `xfm` to MNI standard space. Parameters ---------- subject : str Subject identifier xfm : str Name of functional space transform. Can be 'identity' for anat space. template : str, optional Path to MNI template volume. Defaults to FSL's MNI152_T1_1mm_brain. Returns ------- numpy.ndarray Transformation matrix from the space specified by `xfm` to MNI space. """ # Set up some paths anat_to_mni_xfm = tempfile.mktemp() # Get anatomical image anat_filename = db.get_anat(subject, "brainmask").get_filename() # First use flirt to align masked subject anatomical to MNI template cmd = shlex.split(" ".join(["{fslprefix}flirt".format(fslprefix=fslprefix), "-searchrx -180 180", "-searchry -180 180", "-searchrz -180 180", "-ref", template, "-in", anat_filename, "-omat", anat_to_mni_xfm])) subprocess.call(cmd) # Then load that transform and concatenate it with the functional to anatomical transform anat_to_mni = np.loadtxt(anat_to_mni_xfm) func_to_anat = db.get_xfm(subject, xfm).to_fsl(anat_filename) func_to_mni = np.dot(anat_to_mni, func_to_anat) return func_to_mni
[docs] def transform_to_mni(volumedata, func_to_mni, template=default_template): """ Transform data in `volumedata` to MNI space, resample at the resolution of the atlas image. Parameters ---------- volumedata : VolumeData Data to be transformed to MNI space. func_to_mni : numpy.ndarray Transformation matrix from the space of `volumedata` to MNI space. Get this from `compute_mni_transform`. template : str, optional Path to MNI template volume, used as reference for flirt. Defaults to FSL's MNI152_T1_1mm_brain. Returns ------- nibabel.nifti1.Nifti1Image `volumedata` after transformation to MNI space. """ # Set up paths func_nii = tempfile.mktemp(".nii.gz") func_to_mni_xfm = tempfile.mktemp(".mat") func_in_mni = tempfile.mktemp(".nii.gz") # Save out relevant things volumedata.save_nii(func_nii) _save_fsl_xfm(func_to_mni_xfm, func_to_mni) # Use flirt to resample functional data subprocess.call(["{fslprefix}flirt".format(fslprefix=fslprefix), "-in", func_nii, "-ref", template, "-applyxfm", "-init", func_to_mni_xfm, "-out", func_in_mni]) return nibabel.load(func_in_mni)
[docs] def transform_surface_to_mni(subject, surfname): """ Transform the surface named `surfname` for subject called `subject` into MNI coordinates. Returns [(lpts, lpolys), (rpts, rpolys)]. Parameters ---------- subject : str Subject identifier surfname : str Surface identifier Returns ------- [(mni_lpts, lpolys), (mni_rpts, rpolys)] MNI-transformed surface in same format returned by db.get_surf. """ # Get MNI affine transform mni_affine = nibabel.load(default_template).affine # Get subject anatomical-to-MNI transform mni_xfm = np.dot(mni_affine, db.get_mnixfm(subject, "identity")) # Get transform from surface points to anatomical space ident_xfm = db.get_xfm(subject, "identity", xfmtype="coord") # Get surfaces (lpts, lpolys), (rpts, rpolys) = db.get_surf(subject, surfname) # Transform surface points into anatomical space anat_lpts, anat_rpts = ident_xfm(lpts), ident_xfm(rpts) # Transform anatomical space points to MNI space mni_lpts, mni_rpts = [np.dot(mni_xfm, np.hstack([p, np.ones((p.shape[0],1))]).T).T[:,:3] for p in (anat_lpts, anat_rpts)] return [(mni_lpts, lpolys), (mni_rpts, rpolys)]
[docs] def transform_mni_to_subject(subject, xfm, volarray, func_to_mni, template=default_template): """ Transform data in `volarray` from MNI space to functional space specified by `xfm`. Parameters ---------- subject : str Subject identifier xfm : str Name of functional space that data will be transformed into. volarray : numpy.ndarray 3D volume in MNI space (should have same size as `template`) func_to_mni : numpy.ndarray Transformation matrix from `xfm` space to MNI space. Get this from compute_mni_transform. template : str, optional Path to MNI template volume, used as reference. Defaults to FSL's MNI152_T1_1mm_brain. Returns ------- nibabel.nifti1.Nifti1Image `volarray` after transformation from MNI space to space specified by `xfm`. """ # Set up paths mnispace_func_nii = tempfile.mktemp(".nii.gz") mni_to_func_xfm = tempfile.mktemp(".mat") funcspace_nii = tempfile.mktemp(".nii.gz") # Save out relevant things affine = nibabel.load(template).affine nibabel.save(nibabel.Nifti1Image(volarray, affine), mnispace_func_nii) _save_fsl_xfm(mni_to_func_xfm, np.linalg.inv(func_to_mni)) # Use flirt to resample data to functional space ref_filename = db.get_xfm(subject, xfm).reference.get_filename() subprocess.call(["{fslprefix}flirt".format(fslprefix=fslprefix), "-in", mnispace_func_nii, "-ref", ref_filename, "-applyxfm", "-init", mni_to_func_xfm, "-out", funcspace_nii]) return nibabel.load(funcspace_nii)