[MINC-development] accuracy in mritotal. better tool?

[Peter NEELIN]

On 5 Aug 2003, Yasunari Tosa wrote:

> Even though the rotation and the scaling parts are quite stable, the
> translation part error
> is quite large (up to 1.48 mm).

I'm not quite clear on how you are computing your differences, but you do
need to be careful when comparing transformation matrices. If you simply
take the difference of the translation part of the matrix, you can get
large errors if the centre of the volume is displaced a long way from the
origin (0,0,0), since the rotation is generally being done around the
centre of the brain (roughly). Taking the translation part of the matrix
gives a rotation around the origin, so the translation part can include a
large component arising from rotational and scaling error, even though the
error in the region of the brain is fairly small.

Generally, you need to know where the brain lies in the coordinate system.
At the very least, you need to know what the centre of the brain is. One
way to estimate this is to simply load the volume into register and read
off the initial coordinate. Another way is to find the dimension sizes
with mincinfo and then use voxeltoworld. If, for example, you have a volume
with 128 slices that are 256x256, then use

   voxeltoworld file.mnc 64 128 128

to get the centre of the volume. (A better approach would be to use
something like mincstats to estimate the centroid.) Then use xfm2param:

   xfm2param file.xfm -center <x> <y> <z>

Use the same centre for both xfm files and you should be able to compare
the translations.

(I notice that xfm2param takes a minc file argument - Louis, does that
estimate the centre of gravity or somesuch?)

Anyway, as I said, I am not clear how you are estimating error, so perhaps
you have taken this into account already, or your volumes are already
centred about the origin. If so, then we'll see if Louis (Collins) has any
suggestions as to why the results may be varying.

            Peter
----
            Peter Neelin (neelin@bic.mni.mcgill.ca)