Forked from
FSL / fslpy
521 commits behind the upstream repository.
-
Paul McCarthy authoredPaul McCarthy authored
x5.py 22.40 KiB
#!/usr/bin/env python
#
# x5.py - Functions for working with BIDS X5 files.
#
# Author: Paul McCarthy <pauldmccarthy@gmail.com>
#
"""This module contains functions for reading/writing linear/non-linear
transformations from/to BIDS X5 files. The following functions are available:
.. autosummary::
:nosignatures:
inferType
readLinearX5
writeLinearX5
readNonLinearX5
writeNonLinearX5
.. warning:: This is a development release, and is subject to change.
An X5 file is a HDF5 container file which stores a linear or non-linear
transformation from one NIfTI image to another.
Several terms may be used to refer to these images, such as **source** and
**reference**, **moving** and **fixed**, **from** and **to**, etc. In an X5
file, the two images are simply referred to as **A** and **B**, where **A**
refers to the starting point of the transformation, and **B** to the end
point.
X5 files enable a transformation from the **world coordinate system** of image
**A** to the **world coordinate system** of image **B**. The **world
coordinate system** of an image is defined by its ``sform`` or ``qform``
(hereafter referred to as the ``sform``), which is contained in its NIfTI
header.
Custom HDF5 groups
==================
HDF5 files are composed primarily of *groups*, *attributes*, and
*datasets*:
- *Groups* are simply containers for attributes, datasets, and other groups.
- *Datasets* are strongly-typed, structured N-dimensional arrays.
- *Attributes* are strongly-typed scalar values associated with a group or
dataset.
To simplify the file format definitions below, we shall first define a few
custom HDF5 groups. In the file format definitions, a HDF5 group which is
listed as being of one of these custom types shall contain the
attributes/datasets that are listed here.
*affine*
--------
A HDF5 group which is listed as being of type *affine* contains an affine
transformation, which can be used to transform coordinates from one space into
another. Groups of type *affine* have the following fields:
+-------------+-----------+---------------------------------------------------+
| **Name** | **Type** | **Value/Description** |
+-------------+-----------+---------------------------------------------------+
| ``Type`` | attribute | ``'affine'`` |
+-------------+-----------+---------------------------------------------------+
| ``Matrix`` | dataset | The affine transformation matrix - a ``float64`` |
| | | array of shape ``(4, 4)`` |
+-------------+-----------+---------------------------------------------------+
| ``Inverse`` | dataset | Optional pre-calculated inverse |
+-------------+-----------+---------------------------------------------------+
*space*
-------
A HDF5 group which is listed as being of type *space* contains all of the
information required to define the space of a NIfTI image, including its
shape, dimensions, and voxel-to-world affine transformation.
Groups of type *space* have the following fields:
+--------------+-----------+--------------------------------------------------+
| **Name** | **Type** | **Value/Description** |
+--------------+-----------+--------------------------------------------------+
| ``Type`` | attribute | ``'image'`` |
+--------------+-----------+--------------------------------------------------+
| ``Size`` | attribute | ``uint64`` ``(X, Y, Z)`` voxel dimensions |
+--------------+-----------+--------------------------------------------------+
| ``Scales`` | attribute | ``float64`` ``(X, Y, Z)`` voxel pixdims |
+--------------+-----------+--------------------------------------------------+
| ``Mapping/`` | affine | The image voxel-to-world transformation (its |
| | | ``sform``) |
+--------------+-----------+--------------------------------------------------+
*deformation*
-------------
A HDF5 group which is listed as being of type *deformation* contains a
non-linear transformation, which can be used to transform coordinates from
one space (space **A**) into another (space **B**).
The transformation is represented as a 3D **deformation field** which, at each
voxel within the field, may contain:
- *relative displacements* from space **A** to space **B** (i.e. for a given
location in space **A**, you can add the displacement values to the
coordinates of that location to obtain the coordinates of the corresponding
location in space **B**).
- *absolute coordinates* in space **B**.
The ``Mapping`` affine can be used to calculate a correspondence between the
deformation field coordinate system and the coordinate system of space **A** -
it is assumed that space **A** and the deformation field share a common world
coordinate system.
Groups of type *deformation* have the following fields:
+--------------+-----------+--------------------------------------------------+
| **Name** | **Type** | **Value/Description** |
+--------------+-----------+--------------------------------------------------+
| ``Type`` | attribute | ``'deformation'`` |
+--------------+-----------+--------------------------------------------------+
| ``SubType`` | attribute | ``'absolute'`` or ``'relative'``. |
+--------------+-----------+--------------------------------------------------+
| ``Matrix`` | dataset | The deformation field - a ``float64`` array of |
| | | shape ``(X, Y, Z, 3)`` |
+--------------+-----------+--------------------------------------------------+
| ``Mapping/`` | affine | The field voxel-to-world transformation (its |
| | | ``sform``) |
+--------------+-----------+--------------------------------------------------+
Linear X5 files
===============
Linear X5 transformation files contain an affine transformation matrix of
shape ``(4, 4)``, which can be used to transform image **A** world
coordinates into image **B** world coordinates.
Linear X5 transformation files are assumed to adhere to the HDF5 structure
defined in the table below. All fields are required unless otherwise noted.
+-----------------+-----------+-----------------------------------------------+
| **Name** | **Type** | **Value/Description** |
+-----------------+-----------+-----------------------------------------------+
| *Metadata* |
+-----------------+-----------+-----------------------------------------------+
| ``/Format`` | attribute | ``'X5'`` |
+-----------------+-----------+-----------------------------------------------+
| ``/Version`` | attribute | ``'0.0.1'`` |
+-----------------+-----------+-----------------------------------------------+
| ``/Metadata`` | attribute | JSON string containing unstructured metadata. |
+-----------------+-----------+-----------------------------------------------+
| *Transformation* |
+-----------------+-----------+-----------------------------------------------+
| ``/Type`` | attribute | ``'linear'`` |
+-----------------+-----------+-----------------------------------------------+
| ``/Transform/`` | affine | Affine transformation from image **A** world |
| | | coordinates to image **B** world coordinates |
+-----------------+-----------+-----------------------------------------------+
| ``/A/`` | space | Image **A** space |
+-----------------+-----------+-----------------------------------------------+
| ``/B/`` | space | Image **B** space |
+-----------------+-----------+-----------------------------------------------+
Storage of FSL FLIRT matrices in linear X5 files
------------------------------------------------
FLIRT outputs the result of a linear registration from a source image to a
reference image as an affine matrix of shape ``(4, 4)``. This matrix encodes a
transformation from source image **FSL coordinates** to reference image **FSL
coordinates** [*]_.
In contrast, X5 matrices encode a transformation in **world coordinates**,
i.e. they can be used to transform coordinates from the source image to the
reference image, after both images have been transformed into a common
coordinate system via their respective ``sform`` affines.
The :mod:`fsl.transform.flirt` module contains functions for converting
between FLIRT-style matrices and X5 style matrices.
.. [*] For a given image, FSL coordinates are voxel coordinates scaled by the
``pixdim`` values in the NIFTI header, with an inversion along the X
axis if the voxel-to-world affine (the ``sform``) has a positive
determinant.
Non-linear X5 files
===================
Non-linear X5 transformation files contain a non-linear transformation from
image **A** world coordinates to image **B** world coordinates. The
transformation is represented as a 3D **deformation field** which, at each
voxel within the field, may contain:
- *relative displacements* from image **A** to image **B** (i.e. for a given
location in the image **A** world coordinate system, add the displacement
values to the coordinates to obtain the corresponding location in the
image **B** world coordinate system).
- *absolute coordinates* in the image **B** world coordinate system.
File format specification
-------------------------
Non-linear X5 transformation files are assumed to adhere to the following
HDF5 structure. All fields are required unless otherwise noted.
+---------------+-----------+-------------------------------------------------+
| **Name** | **Type** | **Value/Description** |
+---------------+-----------+-------------------------------------------------+
| *Metadata* |
+---------------+-----------+-------------------------------------------------+
| ``/Format`` | attribute | ``'X5'`` |
+---------------+-----------+-------------------------------------------------+
| ``/Version`` | attribute | ``'0.0.1'`` |
+---------------+-----------+-------------------------------------------------+
| ``/Metadata`` | attribute | JSON string containing unstructured metadata. |
+---------------+-----------+-------------------------------------------------+
| *Transformation* |
+-----------------+-------------+---------------------------------------------+
| **Name** | **Type** | **Value/Description** |
+-----------------+-------------+---------------------------------------------+
| ``/Type`` | attribute | ``'nonlinear'`` |
+-----------------+-------------+---------------------------------------------+
| ``/Transform/`` | deformation | The deformation field, encoding a nonlinear |
| | | transformation from image **A** to image |
| | | **B** |
+-----------------+-------------+---------------------------------------------+
| ``/Inverse/`` | deformation | Optional pre-calculated inverse, encoding a |
| | | nonlinear transformation from image **B** |
| | | to image **A** |
+-----------------+-------------+---------------------------------------------+
| ``/A/`` | space | Image **A** space |
+-----------------+-------------+---------------------------------------------+
| ``/B/`` | space | Image **B** space |
+-----------------+-------------+---------------------------------------------+
Storage of FSL FNIRT warp fields in non-linear X5 files
-------------------------------------------------------
FLIRT outputs the result of a non-linear registration between a source image
and a reference image as either a warp field, or a coefficient field which can
be used to generate a warp field. A warp field is defined in terms of the
reference image - the warp field has the same shape and FOV as the reference
image, and contains either:
- relative displacements from the corresponding reference image location to
the unwarped source image location
- absolute unwarped source image coordinates
The reference image for a FNIRT warp field thus corresponds to image **A** in
a X5 non-linear transform, and the FNIRT source image to image **B**.
FNIRT warp fields are defined in FSL coordinates - a relative warp contains
displacements from reference image FSL coordinates to source image FSL
coordinates, and an absolute warp contains source image FSL coordinates.
When a FNIRT warp field is stored in an X5 file, the displacements/coordinates
must be adjusted so that they encode a transformation from reference image
world coordinates to source image world coordinates.
Conversion of FNIRT coefficient fields
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
A FNIRT coefficient field can be used to generate a deformation field which
contains relative displacements from reference image FSL coordinates to source
image FSL coordinates. If an initial affine registration was used as the
starting point for FNIRT, this generated displacement field contains relative
displacements from the reference image to the *aligned* source image,
i.e. after it has been transformed by the initial affine alignment.
When a FNIRT coefficient field is stored in an X5 file, it must first be
converted to a displacement field. The displacements must then be adjusted so
that they take into account the initial affine alignment (if relevant), and
so that they encode displacements from reference image world coordinates
to source image world coordinates.
The :mod:`fsl.transform.fnirt` module contains functions which can be used to
perform all of the conversions and adjustments required to store FNIRT
transformations as X5 files.
"""
import json
import numpy as np
import nibabel as nib
import h5py
import fsl.version as fslversion
import fsl.data.image as fslimage
from . import affine
from . import nonlinear
X5_FORMAT = 'X5'
X5_VERSION = '0.1.0'
class X5Error(Exception):
"""Error raised if an invalid/incompatible file is detected. """
pass
def inferType(fname):
"""Return the type of the given X5 file - either ``'linear'`` or
``'nonlinear'``.
:arg fname: Name of a X5 file
:returns: ``'linear'`` or ``'nonlinear'``
"""
with h5py.File(fname, 'r') as f:
ftype = f.attrs.get('Type')
if ftype not in ('linear', 'nonlinear'):
raise X5Error('Unknown type: {}'.format(ftype))
return ftype
def readLinearX5(fname):
"""Read a linear X5 transformation file from ``fname``.
:arg fname: File name to read from
:returns: A tuple containing:
- A ``(4, 4)`` ``numpy`` array containing the affine
transformation
- A :class:`.Nifti` instance representing the source space
- A :class:`.Nifti` instance representing the reference space
"""
with h5py.File(fname, 'r') as f:
if f.attrs['Type'] != 'linear':
raise X5Error('Not a linear transform')
_readMetadata( f['/'])
xform = _readAffine(f['/Transform'])
src = _readSpace( f['/A'])
ref = _readSpace( f['/B'])
return xform, src, ref
def writeLinearX5(fname, xform, src, ref):
"""Write a linear transformation to ``fname``.
:arg fname: File name to write to
:arg xform: ``(4, 4)`` ``numpy`` array containing the affine transformation
:arg src: :class:`.Nifti` image representing the source space
:arg ref: :class:`.Nifti` image representing the reference
"""
with h5py.File(fname, 'w') as f:
f.attrs['Type'] = 'linear'
_writeMetadata(f)
_writeAffine( f.create_group('/Transform'), xform)
_writeSpace( f.create_group('/A'), src)
_writeSpace( f.create_group('/B'), ref)
def readNonLinearX5(fname):
"""Read a nonlinear X5 transformation file from ``fname``.
:arg fname: File name to read from
:returns: A :class:`.DeformationField`
"""
with h5py.File(fname, 'r') as f:
if f.attrs.get('Type') != 'nonlinear':
raise X5Error('Not a nonlinear transform')
_readMetadata(f)
ref = _readSpace( f['/A'])
src = _readSpace( f['/B'])
field, xform, defType = _readDeformation(f['/Transform'])
return nonlinear.DeformationField(field,
xform=xform,
src=src,
ref=ref,
srcSpace='world',
refSpace='world',
defType=defType)
def writeNonLinearX5(fname, field):
"""Write a nonlinear X5 transformation to ``fname``.
:arg fname: File name to write to
:arg field: A :class:`.DeformationField`
"""
with h5py.File(fname, 'w') as f:
f.attrs['Type'] = 'nonlinear'
_writeMetadata(f)
_writeSpace( f.create_group('/A'), field.ref)
_writeSpace( f.create_group('/B'), field.src)
_writeDeformation(f.create_group('/Transform'), field)
def _readMetadata(group):
"""Reads a metadata block from the given group, and raises a :exc:`X5Error`
if it does not look valid.
:arg group: A ``h5py.Group`` object.
:returns: A ``dict`` containing the metadata.
"""
format = group.attrs.get('Format')
version = group.attrs.get('Version')
meta = group.attrs.get('Metadata')
parserver = fslversion.parseVersionString(X5_VERSION)
filever = fslversion.parseVersionString(version)
if (format != X5_FORMAT) or (filever[0] != parserver[0]):
raise X5Error('Incompatible format/version (required: {}/{}, '
'present: {}/{})'.format(X5_FORMAT, X5_VERSION,
format, version))
meta = {}
meta['Format'] = format
meta['Version'] = version
meta['Metadata'] = meta
return meta
def _writeMetadata(group):
"""Writes a metadata block into the given group.
:arg group: A ``h5py.Group`` object.
"""
group.attrs['Format'] = X5_FORMAT
group.attrs['Version'] = X5_VERSION
group.attrs['Metadata'] = json.dumps({'fslpy' : fslversion.__version__})
def _readAffine(group):
"""Reads an *affine* from the given group.
:arg group: A ``h5py.Group`` object.
:returns: ``numpy`` array containing a ``(4, 4)`` affine transformation.
"""
if group.attrs.get('Type') != 'affine':
raise X5Error('Not an affine')
xform = group['Matrix']
if xform.shape != (4, 4):
raise X5Error('Not an affine')
return np.array(xform)
def _writeAffine(group, xform):
"""Writes the given affine transformation and its inverse to the given
group.
:arg group: A ``h5py.Group`` object.
:arg xform: ``numpy`` array containing a ``(4, 4)`` affine transformation.
"""
xform = np.asarray(xform, dtype=np.float64)
inv = np.asarray(affine.invert(xform), dtype=np.float64)
group.attrs['Type'] = 'affine'
group.create_dataset('Matrix', data=xform)
group.create_dataset('Inverse', data=inv)
def _readSpace(group):
"""Reads a *space* group, defining a source or reference space.
:arg group: A ``h5py.Group`` object.
:returns: :class:`.Nifti` object. defining the mapping
"""
if group.attrs.get('Type') != 'image':
raise X5Error('Not an image space')
shape = group.attrs['Size']
pixdim = group.attrs['Scales']
xform = _readAffine(group['Mapping'])
hdr = nib.Nifti2Header()
hdr.set_data_shape(shape)
hdr.set_zooms( pixdim)
hdr.set_sform( xform, 'aligned')
return fslimage.Nifti(hdr)
def _writeSpace(group, img):
"""Writes a space specified by ``img`` to the given group.
:arg group: A ``h5py.Group`` object.
:arg img: :class:`.Nifti` object. defining the mapping
"""
group.attrs['Type'] = 'image'
group.attrs['Size'] = np.asarray(img.shape[ :3], np.uint32)
group.attrs['Scales'] = np.asarray(img.pixdim[:3], np.float32)
mapping = group.create_group('Mapping')
_writeAffine(mapping, img.getAffine('voxel', 'world'))
def _readDeformation(group):
"""Reads a *deformation* from the given group.
:arg group: A ``h5py.Group`` object
:returns: A tuple containing
- A ``numpy.array`` containing the deformation field
- A ``numpy.array`` of shape ``(4, 4)`` containing the
voxel to world affine for the deformation field
- The deformation type - either ``'absolute'`` or
``'relative'``
"""
type = group.attrs.get('Type')
subtype = group.attrs['SubType']
if type != 'deformation':
raise X5Error('Not a deformation')
if subtype not in ('absolute', 'relative'):
raise X5Error('Unknown deformation type: {}'.format(subtype))
mapping = _readAffine(group['Mapping'])
field = group['Matrix']
if len(field.shape) != 4 or field.shape[3] != 3:
raise X5Error('Invalid shape for deformation field')
return np.array(field), mapping, subtype
def _writeDeformation(group, field):
"""Write a deformation field to the given group.
:arg group: A ``h5py.Group`` object
:arg field: A :class:`.DeformationField` object
"""
if field.srcSpace != 'world' or \
field.refSpace != 'world':
raise X5Error('Deformation field must encode a '
'world<->world transformation')
group.attrs['Type'] = 'deformation'
group.attrs['SubType'] = field.deformationType
mapping = group.create_group('Mapping')
group.create_dataset('Matrix', data=field.data)
_writeAffine(mapping, field.getAffine('voxel', 'world'))