Commit 30e8c661 authored by Paul McCarthy's avatar Paul McCarthy 🚵
Browse files

RF: little tweaks

parent 48aaf997
%% Cell type:markdown id: tags:
# `fslpy`
**Important:** Portions of this practical require `fslpy` 2.9.0, due to be
released with FSL 6.0.4, in Spring 2020.
[`fslpy`]( is a
Python library which is built into FSL, and contains a range of functionality
for working with FSL and with neuroimaging data from Python.
This practical highlights some of the most useful features provided by
`fslpy`. You may find `fslpy` useful if you are writing Python code to
perform analyses and image processing in conjunction with FSL.
* [The `Image` class, and other data types](#the-image-class-and-other-data-types)
* [Creating images](#creating-images)
* [Working with image data](#working-with-image-data)
* [Loading other file types](#loading-other-file-types)
* [NIfTI coordinate systems](#nifti-coordinate-systems)
* [Transformations and resampling](#transformations-and-resampling)
* [FSL wrapper functions](#fsl-wrapper-functions)
* [In-memory images](#in-memory-images)
* [Loading outputs into Python](#loading-outputs-into-python)
* [The `fslmaths` wrapper](#the-fslmaths-wrapper)
* [The `FileTree`](#the-filetree)
* [Describing your data](#describing-your-data)
* [Using the `FileTree`](#using-the-filetree)
* [Building a processing pipeline with `FileTree`](#building-a-processing-pipeline-with-filetree)
* [The `FileTreeQuery`](#the-filetreequery)
* [Calling shell commands](#calling-shell-commands)
* [The `runfsl` function](#the-runfsl-function)
* [Submitting to the cluster](#submitting-to-the-cluster)
* [Redirecting output](#redirecting-output)
* [FSL atlases](#fsl-atlases)
* [Querying atlases](#querying-atlases)
* [Loading atlas images](#loading-atlas-images)
* [Working with atlases](#working-with-atlases)
> **Note**: `fslpy` is distinct from `fslpython` - `fslpython` is the Python
> environment that is baked into FSL. `fslpy` is a Python library which is
> installed into the `fslpython` environment.
Let's start with some standard imports and environment set-up:
%% Cell type:code id: tags:
%matplotlib inline
import matplotlib.pyplot as plt
import os
import os.path as op
import nibabel as nib
import numpy as np
import warnings
np.set_printoptions(suppress=True, precision=4)
%% Cell type:markdown id: tags:
And a little function that we can use to generate a simple orthographic plot:
%% Cell type:code id: tags:
def ortho(data, voxel, fig=None, cursor=False, **kwargs):
"""Simple orthographic plot of a 3D array using matplotlib.
:arg data: 3D numpy array
:arg voxel: XYZ coordinates for each slice
:arg fig: Existing figure and axes for overlay plotting
:arg cursor: Show a cursor at the `voxel`
All other arguments are passed through to the `imshow` function.
:returns: The figure and orthogaxes (which can be passed back in as the
`fig` argument to plot overlays).
voxel = [int(round(v)) for v in voxel]
data = np.asanyarray(data, dtype=np.float)
data[data <= 0] = np.nan
x, y, z = voxel
xslice = np.flipud(data[x, :, :].T)
yslice = np.flipud(data[:, y, :].T)
zslice = np.flipud(data[:, :, z].T)
if fig is None:
fig = plt.figure()
xax = fig.add_subplot(1, 3, 1)
yax = fig.add_subplot(1, 3, 2)
zax = fig.add_subplot(1, 3, 3)
fig, xax, yax, zax = fig
xax.imshow(xslice, **kwargs)
yax.imshow(yslice, **kwargs)
zax.imshow(zslice, **kwargs)
if cursor:
cargs = {'color' : (0, 1, 0), 'linewidth' : 1}
xax.axvline( y, **cargs)
xax.axhline(data.shape[2] - z, **cargs)
yax.axvline( x, **cargs)
yax.axhline(data.shape[2] - z, **cargs)
zax.axvline( x, **cargs)
zax.axhline(data.shape[1] - y, **cargs)
for ax in (xax, yax, zax):
return (fig, xax, yax, zax)
%% Cell type:markdown id: tags:
And another function which uses FSLeyes for more complex plots:
%% Cell type:code id: tags:
def render(cmdline):
import shlex
import IPython.display as display
prefix = '-of screenshot.png -hl -c 2 '
from fsleyes.render import main
main(shlex.split(prefix + cmdline))
except ImportError:
# fall-back for macOS - we have to run
# FSLeyes render in a separate process
from import runfsl
prefix = 'render ' + prefix
runfsl(prefix + cmdline, env={})
return display.Image('screenshot.png')
%% Cell type:markdown id: tags:
<a class="anchor" id="the-image-class-and-other-data-types"></a>
## The `Image` class, and other data types
module provides the `Image` class, which sits on top of `nibabel` and contains
some handy functionality if you need to work with coordinate transformations,
or do some FSL-specific processing. The `Image` class provides features such
- Support for NIFTI1, NIFTI2, and ANALYZE image files
- Access to affine transformations between the voxel, FSL and world coordinate
- Ability to load metadata from BIDS sidecar files
> The `Image` class behaves differently to the `nibabel.Nifti1Image`. For
> example, when you create an `Image` object, the default behaviour is to load
> the image data into memory. This is configurable however; take a look at
> [the
> documentation](
> to explore all of the options.
Some simple image processing routines are also provided - these are covered
<a class="anchor" id="creating-images"></a>
### Creating images
It's easy to create an `Image` - you can create one from a file name:
%% Cell type:code id: tags:
from import Image
stddir = op.expandvars('${FSLDIR}/data/standard/')
# load a FSL image - the file
# suffix is optional, just like
# in real FSL-land!
std1mm = Image(op.join(stddir, 'MNI152_T1_1mm'))
%% Cell type:markdown id: tags:
You can create an `Image` from an existing `nibabel` image:
%% Cell type:code id: tags:
# load a nibabel image, and
# convert it into an FSL image
nibimg = nib.load(op.join(stddir, 'MNI152_T1_1mm.nii.gz'))
std1mm = Image(nibimg)
%% Cell type:markdown id: tags:
Or you can create an `Image` from a `numpy` array:
%% Cell type:code id: tags:
data = np.zeros((182, 218, 182))
img = Image(data, xform=np.eye(4))
%% Cell type:markdown id: tags:
If you have generated some data from another `Image` (or from a
`nibabel.Nifti1Image`) you can use the `header` option to set
the header information on the new image:
%% Cell type:code id: tags:
img = Image(data, header=std1mm.header)
%% Cell type:markdown id: tags:
You can save an image to file via the `save` method:
%% Cell type:code id: tags:
%% Cell type:markdown id: tags:
`Image` objects have all of the attributes you might expect:
%% Cell type:code id: tags:
stddir = op.expandvars('${FSLDIR}/data/standard/')
std1mm = Image(op.join(stddir, 'MNI152_T1_1mm'))
print('name: ',
print('file: ', std1mm.dataSource)
print('NIfTI version:', std1mm.niftiVersion)
print('ndim: ', std1mm.ndim)
print('shape: ', std1mm.shape)
print('dtype: ', std1mm.dtype)
print('nvals: ', std1mm.nvals)
print('pixdim: ', std1mm.pixdim)
%% Cell type:markdown id: tags:
and a number of useful methods:
%% Cell type:code id: tags:
std2mm = Image(op.join(stddir, 'MNI152_T1_2mm'))
mask2mm = Image(op.join(stddir, 'MNI152_T1_2mm_brain_mask'))
print(std2mm.getAffine('voxel', 'world'))
%% Cell type:markdown id: tags:
An `Image` object is a high-level wrapper around a `nibabel` image object -
you can always work directly with the `nibabel` object via the `nibImage`
%% Cell type:code id: tags:
%% Cell type:markdown id: tags:
<a class="anchor" id="working-with-image-data"></a>
### Working with image data
You can get the image data as a `numpy` array via the `data` attribute:
%% Cell type:code id: tags:
data =
print(data.min(), data.max())
ortho(data, (45, 54, 45))
%% Cell type:markdown id: tags:
> Note that `` will give you the data in its underlying type, unlike
> the `nibabel.get_fdata` method, which up-casts image data to floating-point.
You can also read and write data directly via the `Image` object:
%% Cell type:code id: tags:
slc = std2mm[:, :, 45]
std2mm[0:10, :, :] *= 2
%% Cell type:markdown id: tags:
Doing so has some advantages that may or may not be useful, depending on your
- The image data will be kept on disk - only the parts that you access will
be loaded into RAM (you will also need to pass`loadData=False` when creating
the `Image` to achieve this).
- The `Image` object will keep track of modifications to the data - this can
be queried via the `saveState` attribute.
<a class="anchor" id="loading-other-file-types"></a>
### Loading other file types
package has a number of other classes for working with different types of FSL
and neuroimaging data. Most of these are higher-level wrappers around the
corresponding `nibabel` types:
* The
class can be used to load a bitmap image (e.g. `jpg, `png`, etc) and
class can be used to load a bitmap image (e.g. `jpg`, `png`, etc) and
convert it to a NIfTI image.
* The
class uses `dcm2niix` to load NIfTI images contained within a DICOM
* The
class can be used too load `.mgh`/`.mgz` images (they are converted into
NIfTI images).
* The
module contains functions for loading and working with the output of the
FSL `dtifit` tool.
* The
modules contain classes and functions for loading data from FEAT
* Similarly, the
modules contain classes and functions for loading data from MELODIC
* The
modules contain functionality form loading surface data from GIfTI,
freesurfer, and ASCII VTK files respectively.
> <sup>*</sup>You must make sure that
> [`dcm2niix`]( is installed on your
> system in order to use this class.
<a class="anchor" id="nifti-coordinate-systems"></a>
### NIfTI coordinate systems
The `Image.getAffine` method gives you access to affine transformations which
can be used to convert coordinates between the different coordinate systems
associated with a NIfTI image. Have some MNI coordinates you'd like to convert
to voxels? Easy!
%% Cell type:code id: tags:
stddir = op.expandvars('${FSLDIR}/data/standard/')
std2mm = Image(op.join(stddir, 'MNI152_T1_2mm'))
mnicoords = np.array([[0, 0, 0],
[0, -18, 18]])
world2vox = std2mm.getAffine('world', 'voxel')
vox2world = std2mm.getAffine('voxel', 'world')
# Apply the world->voxel
# affine to the coordinates
voxcoords = ([:3, :3], mnicoords.T)).T + world2vox[:3, 3]
%% Cell type:markdown id: tags:
The code above is a bit fiddly, so instead of figuring it out, you can just
use the
%% Cell type:code id: tags:
from fsl.transform.affine import transform
voxcoords = transform(mnicoords, world2vox)
# just to double check, let's transform
# those voxel coordinates back into world
# coordinates
backtomni = transform(voxcoords, vox2world)
for m, v, b in zip(mnicoords, voxcoords, backtomni):
print(m, '->', v, '->', b)
%% Cell type:markdown id: tags:
> The `Image.getAffine` method can give you transformation matrices
> between any of these coordinate systems:
> - `'voxel'`: Image data voxel coordinates
> - `'world'`: mm coordinates, defined by the sform/qform of an image
> - `'fsl'`: The FSL coordinate system, used internally by many FSL tools
> (e.g. FLIRT)
Oh, that example was too easy I hear you say? Try this one on for size. Let's
say we have run FEAT on some task fMRI data, and want to get the MNI
coordinates of the voxel with peak activation.
> This is what people used to use `Featquery` for, back in the un-enlightened
> days.
Let's start by identifying the voxel with the biggest t-statistic:
%% Cell type:code id: tags:
featdir = op.join('08_fslpy', 'fmri.feat')
tstat1 = Image(op.join(featdir, 'stats', 'tstat1')).data
# Recall from the numpy practical that
# argmax gives us a 1D index into a
# flattened view of the array. We can
# use the unravel_index function to
# convert it into a 3D index.
peakvox = np.abs(tstat1).argmax()
peakvox = np.unravel_index(peakvox, tstat1.shape)
print('Peak voxel coordinates for tstat1:', peakvox, tstat1[peakvox])
%% Cell type:markdown id: tags:
Now that we've got the voxel coordinates in functional space, we need to
transform them into MNI space. FEAT provides a transformation which goes
directly from functional to standard space, in the `reg` directory:
%% Cell type:code id: tags:
func2std = np.loadtxt(op.join(featdir, 'reg', 'example_func2standard.mat'))
%% Cell type:markdown id: tags:
But ... wait a minute ... this is a FLIRT matrix! We can't just plug voxel
coordinates into a FLIRT matrix and expect to get sensible results, because
FLIRT works in an internal FSL coordinate system, which is not quite
`'voxel'`, and not quite `'world'`. So we need to do a little more work.
Let's start by loading our functional image, and the MNI152 template (the
source and reference images of our FLIRT matrix):
%% Cell type:code id: tags:
func = Image(op.join(featdir, 'reg', 'example_func'))
std = Image(op.expandvars(op.join('$FSLDIR', 'data', 'standard', 'MNI152_T1_2mm')))
%% Cell type:markdown id: tags:
Now we can use them to get affines which convert between all of the different
coordinate systems - we're going to combine them into a single uber-affine,
which transforms our functional-space voxels into MNI world coordinates via:
1. functional voxels -> FLIRT source space
2. FLIRT source space -> FLIRT reference space
3. FLIRT referece space -> MNI world coordinates
%% Cell type:code id: tags:
vox2fsl = func.getAffine('voxel', 'fsl')
fsl2mni = std .getAffine('fsl', 'world')
%% Cell type:markdown id: tags:
Combining two affines into one is just a simple dot-product. There is a
`concat()` function which does this for us, for any number of affines:
%% Cell type:code id: tags:
from fsl.transform.affine import concat
# To combine affines together, we
# have to list them in reverse -
# linear algebra is *weird*.
funcvox2mni = concat(fsl2mni, func2std, vox2fsl)
%% Cell type:markdown id: tags:
> In the next section we will use the
> [`fsl.transform.flirt.fromFlirt`](
> function, which does all of the above for us.
So we've now got some voxel coordinates from our functional data, and an
affine to transform into MNI world coordinates. The rest is easy:
%% Cell type:code id: tags:
mnicoords = transform(peakvox, funcvox2mni)
mnivoxels = transform(mnicoords, std.getAffine('world', 'voxel'))
mnivoxels = [int(round(v)) for v in mnivoxels]
print('Peak activation (MNI coordinates):', mnicoords)
print('Peak activation (MNI voxels): ', mnivoxels)
%% Cell type:markdown id: tags:
Note that in the above example we are only applying a linear transformation
into MNI space - in reality you would also want to apply your non-linear
structural-to-standard transformation too. This is covered in the next
<a class="anchor" id="transformations-and-resampling"></a>
### Transformations and resampling
Now, it's all well and good to look at t-statistic values and voxel
coordinates and so on and so forth, but let's spice things up a bit and look
at some images. Let's display our peak activation location in MNI space. To do
this, we're going to resample our functional image into MNI space, so we can
overlay it on the MNI template. This can be done using some handy functions
from the
Let's make sure we've got our source and reference images loaded:
%% Cell type:code id: tags:
featdir = op.join(op.join('08_fslpy', 'fmri.feat'))
tstat1 = Image(op.join(featdir, 'stats', 'tstat1'))
std = Image(op.expandvars(op.join('$FSLDIR', 'data', 'standard', 'MNI152_T1_2mm')))
%% Cell type:markdown id: tags:
Now we'll load the `example_func2standard` FLIRT matrix, and adjust it so that
it transforms from functional *world* coordinates into standard *world*
coordinates - this is what is expected by the `resampleToReference` function,
used below:
%% Cell type:code id: tags:
from fsl.transform.flirt import fromFlirt
func2std = np.loadtxt(op.join(featdir, 'reg', 'example_func2standard.mat'))
func2std = fromFlirt(func2std, tstat1, std, 'world', 'world')
%% Cell type:markdown id: tags:
Now we can use `resampleToReference` to resample our functional data into
MNI152 space. This function returns a `numpy` array containing the resampled
data, and an adjusted voxel-to-world affine transformation. But in this case,
we know that the data will be aligned to MNI152, so we can ignore the affine:
%% Cell type:code id: tags:
from fsl.utils.image.resample import resampleToReference
std_tstat1 = resampleToReference(tstat1, std, func2std)[0]
std_tstat1 = Image(std_tstat1, header=std.header)
%% Cell type:markdown id: tags:
Now that we have our t-statistic image in MNI152 space, we can plot it in
standard space using `matplotlib`:
%% Cell type:code id: tags:
stddir = op.expandvars('${FSLDIR}/data/standard/')
std2mm = Image(op.join(stddir, 'MNI152_T1_2mm'))
std_tstat1 =
std_tstat1[std_tstat1 < 3] = 0
fig = ortho(, mnivoxels,
fig = ortho(std_tstat1, mnivoxels,, fig=fig, cursor=True)
%% Cell type:markdown id: tags:
In the example above, we resampled some data from functional space into
standard space using a linear transformation. But we all know that this is not
how things work in the real world - linear transformations are for kids. The
real world is full of lions and tigers and bears and warp fields.