Anatomist uses AIMS to handle its referentials so behaves exactly the same way.

Aims tries to work internally in an image-specific referential, but with always the same orientation. This orientation is axial with the following coordinates system:

Once loaded in memory, all voxels should be organized in this order. As a consequence, images in Anatomist are always displayed in radiological mode, whatever the actual orientation of data on disk.

Images on disk, depending on their format and acquisition modes, are not necessarily in this orientation. When a different orientation is detected, images are flipped in memory at load-time to fit the standard AIMS orientation. And when images are written back to disk, they may also be flipped back according to the specific format needs.

By default, AIMS doesn't apply any transformation other than flipping images at load time as described just before.

But transformations can be provided in some Aims commands or loaded in Anatomist to apply coordinates changes. Then coords transformations are applied on the fly when processing or displaying data which are not in the same referential.

There is no special referential (such as a common central working referential).

Transformation files used by AIMS (.trm files) are ASCII files looking like this:

Tx Ty Tz
R11 R12 R13
R21 R22 R23
R31 R32 R33

    [ R11  R12  R13  Tx ]
M = [ R21  R22  R23  Ty ]
    [ R31  R32  R33  Tz ]
    [   0    0    0   1 ]

AIMS (and Anatomist) writes an additional header file which can store any additional information: the .minf header (for Meta-INFormation) when saving its data (images, meshes, and any other data), and reads it if it is present when loading data files. This meta-header has the shape displayed by the AimsFileInfo command, and may be saved in "python dictionary" or XML formats. The MINF file has the same file name as the main data file, with the .minf extension added (toto.img.minf for instance).

The MINF header may contain referentials and transformations information. When present, this information is stored in a few fields:

attributes = {
  'storage_to_memory' : [ 1, 0, 0, 0, 0, -1, 0, 62, 0, 0, -1, 45, 0, 0, 0, 1 ],
  'referentials' : [ 'Coordinates aligned to another file or to anatomical truth' ],
  'transformations' : [ [ -1, 0, 0, 78, 0, -1, 0, 75, 0, 0, -1, 84, 0, 0, 0, 1 ] ],
  'referential': 'be9724cc-eceb-d831-a83e-335e12b80f14',
}

Internally, SPM thinks things are always in the same orientation, which is also axial but with different axes:

Working on the coordinate transformations for years and regularly getting headaches from it, I am still not 100% sure of what I say here, so if I'm wrong, please correct me by sending a message on BrainVisa forum (http://brainvisa.info/forum/). Especially, I'm not sure that SPM99 and SPM2 really use the same referentials.

SPM handles input Analyze images in two different orientations: axial radiological and axial neurological orientations. This orientation is not specified in SPM-Analyze format image files, so you have to tell how they are oriented. This is done in SPM by a flipping flag set somewhere in SPM defaults configuration (default.analyze.flip in SPM2).

This is specific to SPM-Analyze format, and does not apply to NIFTI-1 or Minc formats. Hopefully the Analyze format is now obsolete and will disapear with time, but there are still existing files...

This flipping flag has changed in form and meaning between SPM99 and SPM2.

As I have understood:

SPM uses a common central referential to work in. Every image can provide a transformation matrix to this referential. Such a transformation may be specified in different ways:

Normalization files (*_sn3d.mat for SPM99, *_sn.mat for SPM2 and newer) contain transformations to the referential of a normalization template (either a standard one provided with the SPM software distribution, or a custom user-made one). This transformation contains an affine part (matrix), and optionally, depending on the normalization type, a non-linear part (coefficients on a functions base as far as I know but I don't know much about this part). Information about the input and template images are also included (dimensions, and origins or voxels-to-template transformation).

Normalized images are in the referential of the normalization template used, but not necessarily with the same bounding box, resolution and field of view.

SPM99 and SPM2 use normalization files with different names and different contents. They are not compatible, even if there is some common and similar information in them.

SPM normalization files are in matlab (.mat) format. AIMS cannot read the proprietary matlab format, so such files cannot be directly imported in AIMS.

However, the scipy module for Python language can read them. So we have made Python scripts in PyAims and in BrainVisa to convert SPM matrices to AIMS .trm format. Only the affine part can be converted, because AIMS only use matrices for transformations, and non-linear information cannot fit into a matrix. Look at the AimsSpmNormalizationConvert.py program, and the SPMsn3dToAims process in BrainVISA.

As the orientation is different in SPM and AIMS, a transformation to a template image is not the same as a transformation to a normalized image with a different field of view. So, when converting SPM normalization matrices, the normalized image must be also provided, otherwise BrainVisa can only give the transformation to the normalization template. Note the difference.

It is impossible to guess the orientation of such images if you don't know how they were acquired. This means you have to manually specify their orientation, either for all images in SPM, or in BrainVisa when importing them into a database. BrainVisa tags them so it knows everything afterwards and avoids mistakes. SPM does not.

Normalizing the same image in radiological orientation with SPM99 and SPM2 results in normalized images in different orientations. When you import normalized images coming from another site, you have to know whether they have been normalized by SPM99 or by SPM2, and if the original image was in radiological or neurological orientation.

I think the normalization file (*_sn3d.mat for SPM99, or *_sn.mat for SPM2) contains enough information to retreive the orientation of input and template images, so can disambiguate the situation.

The origin field of SPM format is the position of the referential origin, in voxels and starting from 1 (not from 0). In fact it's a matlab array index. So it is given in the orientation of the image on disk. AIMS flips SPM images on several axes when loading them, so the origin information also has to be flipped. Flipping it needs to know the image dimensions.

AIMS referentials have their origin in the first voxel, (almost) in the corner of the image, and normally don't use the SPM origin. But the origin information is read and maintained. Anatomist can, if asked for, make a transformation going from AIMS origin (corner) to the SPM origin. This allows to display several aligned SPM images in Anatomist with the correct correspondance. However after this translation, the coordinates are still in AIMS orientation (radiological and indirect), not in SPM, so the coordinates do not correspond to what they are in SPM.

To compare coordinates of SPM images in Anatomist and SPM, another transformation has to be applied in Anatomist, with all the flips included. Anatomist can directly use the SPM/MNI normalization referential.

Up to now, GIS images are considered being always in AIMS orientation unless specified in their AIMS meta-header (.minf file, see the corresponding paragraph). No flips are applied.

The Minc and NIFTI-1 IO plugins take orientation into account and flip data accordingly when reading / writing files.

I am not sure if other formats (Dicom, Ecat...) can specify an image orientation or not. If they do, the current release of AIMS will probably not take it into account.

SPM99 and SPM2 don't use the same format of normalization files, but both provide more or less the following information:

We want first A2T, then AIMS, provided Affine, S2MA and MT

  A2T = MT * ( VsA * S2MA * At * Affine )^-1
  AIMS = TN^-1 * A2T