[Mrtrix-discussion] Crossing-fibers gray matter CSD

[Donald Tournier]

Hi Wim,

> - In which voxels do I have to estimate the fiber response function?
> In all brain-voxels (resulting in a less 'flat' response function) or
> in the white matter voxels? As we need it to track in both white
> matter and gray matter voxels.

This is an interesting question, and not one I have a satisfactory
answer to. Mind you, I don't think anyone has a good answer to this.
The few people who have suggested the possibility of tracking in grey
matter (I can only think of Van Wedeen, and maybe indirectly Tim
Behrens for deep grey matter structures) have both used methods that
don't need an explicit response function (diffusion spectrum imaging
and the diffusion tensor, respectively).

Thankfully, CSD is not overly sensitive to inaccuracies in the
response function. I'd recommend you go for the 'flattest' response
function you can get, since this will always produce sharper
directions. This means that you should opt for the white matter
response function. Besides, that is your only real option anyway,
since it won't be possible to isolate even a single voxel within the
grey matter that contains a single coherent fibre direction, from
which you might have been able to estimate a response function...


> - How does streamtrack determine the principal tracking direction(s)
> from the spherical decomposition data? Is it using a 'find_SH_peaks'
> internally? Does it take crossing-fibers into account, or only the
> first direction?

There are various algorithms within streamtrack, but all the SD based
ones will take crossing fibres into account. The SD_STREAM option will
do find the closest peak to the current direction of tracking, and use
its direction for the next step. The SD_PROB option on the other hand,
randomly samples a direction from the current distribution of fibre
orientations (the FOD), within a 'cone' about the current direction of
tracking. The angle of the cone is determined from the curvature
constraint, and is given by phi = 2 asin (s/2R), where s is the step
size and R is the radius of curvature. This means both algorithms will
preferentially track through crossing fibre regions, providing of
course that there is a fibre orientation in that direction.


> - tracks2prob flips my tracks in the x and y plane ( I have to
> 'correct' with fslswapdim -x -y z <input> <output> to get it right
> with the 'reference image'). The fibers are oriented correctly in
> mrview. Am I doing something wrong? (q-form, s-form thing?).

I'm surprised by this. Can you provide more details on what you mean?
Is the resulting image flipped within MRView, or within FSL? As far as
I can tell, FSLview is not very flexible when it comes to data
ordering, and will for example display images acquired in the sagittal
plane 'as if' they had been acquired axial, which will of course look
wrong (although the L-R, A-P, and I-S orientation labels are in the
correct places). I'd imagine that this would also mean that overlaying
two images where the data are ordered differently will produce
artefacts like you mention.

If the problem is with FSL's display, then you may have no other
option than what you've already done (although mrconvert does have a
"-layout" option that could be used to the same effect, and probably
more robustly). If the problem is with MRView, let me know and I will
investigate further.

Hope this helps.
Cheers!

Donald.


--
Jacques-Donald Tournier (PhD)
Brain Research Institute, Melbourne, Australia
Tel: +61 (0)3 9496 4078