.. _image-orientation:

===================
 Image orientation
===================

Every image in ``nibabel`` has an orientation.  The orientation is the
relationship between the voxels in the image array, and millimeters in
some space.  

Affines as orientation
----------------------

Orientations are expressed by 4 by 4 affine arrays.  4x4 affine arrays
give, in homogenous coordinates, the relationship between the
coordinates in the voxel array, and millimeters.  Let is say that I have
a simple affine like this:

>>> import numpy as np
>>> aff = np.diag((2, 3, 4, 1))
>>> aff[:3,3] = [10, 11, 12]

And I have a voxel coordinate:

>>> coord = np.array([3, 2, 1])

then the millimeter coordinate for that voxel is given by:

>>> # add extra 1 for homogenous coordinates
>>> homogenous_coord = np.concatenate((coord, [1]))
>>> mm_coord = np.dot(aff, homogenous_coord)[:3]
>>> mm_coord
array([16, 17, 16])

Affines and image formats
-------------------------

Some image formats (such as nifti) allow storage of affine or
affine-like image orientation, and some do not (such as Analyze).  Almost
all image formats allow you to save an image without any affine
information.  Most image orientation problems arise for images that do
not have full affine information, and we have to guess.

Making an affine when there is no stored affine
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

If there is no affine information, we have to make some best-guess
affine for the image. In this case, the image is assumed to be saved
(fastest to slowest changing) in X, Y, Z dimension order, and we
construct a 4x4 affine ``aff`` where ``aff[:3,:3]`` is a diagonal matrix
with the X, Y, Z zooms (voxel sizes) as entries - ``diag(aff)``.  The
translation part of the affine ``aff[:3, 3]`` is such that the central
voxel in the image is at 0, 0, 0 mm (this is not completely true for SPM
images, with may have encoded a particular voxel as the origiin using
the ``origin`` field of the SPM version of the Analyze header).

The left-right orientation of the image in this case boils down to
whether the first voxel in the image (and in any x line) is the
left-most voxel or the right-most voxel.   If it is the left-most, the
image is said to be in 'neurological' orientation, and if it is the
right-most, it's in 'radiological' orientation.   These terms only
make sense in this case, where there is no affine, and we are assuming
X, Y, Z data storage on disk.

If we deem the image to be 'neurological' then the guessed affine
above will be correct, as a transform from voxel coordinates to mm
coordinates.  If it is 'radiological', then we need to multiply the
'X' zoom (``aff[0,0]``) by -1, and adjust the X translation
(``aff[0,3]``) accordingly.

In ``nibabel`` we assume that any image without an affine has been
stored in radiological order on disk - and thus the guessed affine needs
a left-right flip.  This is true for all Analyze-type image formats
(Analyze, SPM analyze, nifti).

If you want to change this (please don't unless you are absolutely
sure what you are doing), the default is encoded in the
``default_x_flip`` class variable where True corresponds to
'radiological' and False corresponds to 'neurological'.

If you want to load images that are in neurological disk format, I
strongly suggest that, instead of changing this default, you adjust
the affine after loading, as in::

    img = nibabel.load('some_image.img')
    aff = img.get_affine()
    x_flipper = np.diag([-1,1,1,1])
    lr_img = nibabel.Nifti1Image(img.get_data, np.dot(x_flipper, aff), img.get_header())

Affines for Analyze, SPM analyze, and NIFTI
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

Analyze images can't have an affine, so the above always applies to
Analyze images.

SPM99 images are unpleasantly confusing, because they may have an
affine stored in a 'some_image.mat' file, in a matrix called 'M', but the
affine for the image, is given by (from the code above)
``np.dot(x_flipper, M)`` - that is - the affine gives the
transformations to be applied before any left-right flipping, where
left-right flipping is determined by the ``default_x_flip`` above.
Horrible.

SPM2 images are a bit more straightforward, in that there may be an
affine, again stored in the 'some_image.mat' file, but, if the image has
been written in SPM2, or by us, in SPM2 format, then there should be a
'mat' matrix in that file, that has the full affine, which is
unaffected by the ``default_x_flip``.   However, if we are loading
what appears to be an SPM99 image, that only has a mat file with an
'M' matrix, we apply the default flip as above.

Whenever we save an SPM99 image, we save an SPM2-like ``.mat`` file, with
both the flip-specifying 'mat' matrix, and the pre-flip 'M' matrix,
because this is still backwards compatible, and might be less liable
to chaos if someone changes the default flip setting.

Then, we have nifti, which can store two affines, the ``qform`` and
the ``sform``.  If the ``sform`` is present, we load that, otherwise,
if the ``qform`` is present, we use that. Either of these affines
fully specifies orientation, that is, they ignore any settings of
``default_x_flip``.   If the nifti has neither a ``qform`` or an
``sform``, we guess at the affine with the algorithm above, and the
``default_x_flip`` comes into play again.

Note that, for nifti images without affines, we don't followw the nifti
standard.  In the nifti standard, if an image does not have an affine,
then the affine is deemed to be ``diag([xs, ys, zs, 1])`` where ``xs,
ys, zs`` are the X, Y and Z zooms (voxel sizes) respectively.  This
array has no concept of left-right flipping corresponding to
radiological orientation, and assumes the image origin (voxel
corresponding to 0, 0, 0 in millimeters) is the first voxel in the
image.  ``nibabel`` differs from the nifti standard, for images without
affines, in using the center of the image as the origin, and flipping
left-right by default.  We chose this break from the standard because
that is what SPM does with non-affine niftis, and because it seemed more
sensible, and because it's more consistent with what we do with SPM
non-nifti images (not surprisingly).

