[Mrtrix-discussion] tckgen: effect of signal amplitude and streamline length distribution

[Robert Smith]

Hi Ulrike (and all),

a) Under standard usage scenarios, the size of the ODF glyphs shouldn't
vary massively. This is because if the amplitude of the raw DWI amplitude
is scaled up or down, the estimated response function will scale
accordingly, such that deconvolution of a single-fibre voxel should always
give an integral of 1.0 or thereabouts. So there would need to be something
unusual about the derivation of your response function, or scaling of your
image subsequent to RF estimation, for CSD to produce a result like that.
The only case I can think of where large ODFs come naturally is if you've
used tckmap with the -tod option; those need to be scaled explicitly. You
could scale your SH images manually using mrcalc before opening them in the
viewer, but personally I'd go back to the raw data and figure out where the
scaling discrepancy is coming from.

The primary effect that the scaling will have is the default 0.1 threshold
for tracking. Having massive FOD amplitudes is comparable to setting that
threshold to zero; so you'll get a fraction more streamline orientation
dispersion, and streamlines will be able to track small noisy FOD lobes. I
would suggest that this is probably what's causing those inferior-superior
streamlines in the CC. It won't have a tremendous effect on SIFT
(everything will just scale accordingly), though it will affect the
thresholds used during the FOD segmentation process, so noisy FOD lobes
will get included in processing.

b) Looks like you've reproduced my result
<http://www.sciencedirect.com/science/article/pii/S1053811914008155> quite
nicely. There's a lot of reasons why this seems an entirely plausible
attribute of the white matter to me. Long connections give less bandwidth
per unit volume than short ones, so for a fixed volume (like the brain) it
makes more sense to have lots of short connections. Or if two distant
regions require a structural connection, having that connection perform a
'stop-over' along the way provides increased brain connectivity complexity
with little metabolic penalty. Indeed in the reference
<http://sirl.stanford.edu/newlm/images/9/91/SchuzBraitenberg.pdf> I used to
draw the post-mortem comparison in that paper, they light-heartedly
compared this property to the Zipf Law
<http://en.wikipedia.org/wiki/Zipf%27s_law> of linguistics. Although the
dispersion of probabilistic streamlines will contribute to a bias toward
shorter connections, I doubt it's the major factor at play here. I also did
everything I could to ensure that SIFT does not have an internal bias
toward retaining or rejecting streamlines of any particular length; it
comes purely from the reconstruction and the image data.

So next time you see somebody using 25mm as a tracking minimum length, show
them that histogram!

(P.S. If you're using ACT, and doing whole-brain tracking, there's no need
to provide a tracking mask using the -mask option; in fact you're better
off not, in case the mask causes streamlines to terminate before they
ideally should)

Rob


--

*Robert Smith, Ph.D*
Research Officer, Imaging Division

The Florey Institute of Neuroscience and Mental Health
Melbourne Brain Centre - Austin Campus
245 Burgundy Street
Heidelberg Vic 3084
Ph: +61 3 9035 7128
Fax: +61 3 9035 7301
www.florey.edu.au

On Wed, Jan 21, 2015 at 8:02 PM, Ulrike Kuhl <kuhl at cbs.mpg.de> wrote:

> Dear MRTrix team,
>
> I use MRTrix3 to generate streamlines and SIFT for filtering the result.
> I noticed two things where I am not sure if the results are right.
>
> a) I noted that my ODF glyphs appear to be huge when I load them into
> MRview (glyphs not being normalized to the b0 image, just as you propose).
> Even if I completely scroll out of the little ODF display window I am not
> able to see the corpus callosum glyphs completely. Likewise, I have to
> scale the overlay by at least 0.01 in order to distinguish individual
> glyphs. Is that normal?
> If not, might it be ascribable to the fact that my data contains larger
> signal amplitudes than yours (you using TrioTim data with b = 3000; me
> using TrioTim data with b = 1000)? If so, do you see any problems that
> might arise when I use such huge glyphs for streamline generation and
> SIFTing?
>
> b) Using my huge glyphs, I produce 10,000,000 streamlines with tckgen in
> the following way:
>
> tckgen -algorithm iFOD2 -step 1.25 -samples 4 -maxlength 250 -number
> 10000000 -angle 45 -backtrack -force -seed_gmwmi GM_WM_interface.nii.gz
> -mask WM_mask.nii.gz -act 5tt.mif CSD.nii.gz out.tck
>
> Plotting the histogram of resulting streamline lengths shows that by far
> the greatest proportion of these streamlines are really short (I attached
> histograms of streamline lengths before and after SIFT if you want to check
> out the distributions).
> I go then on using SIFT to reduce the number of streamlines to 1,000,000 -
> also here, the greatest proportion are very short fibres. By visual
> inspection I noticed that there remains quite a number of short fibres in
> superior-inferior direction within the corpus callosum. That seems really
> wrong... does that correspond to the distribution of streamline lengths and
> location on your data after SIFT?
>
> Thank you very much for your help,
>
> Ulrike
>
> --
> Max Planck Institute for Human Cognitive and Brain Sciences
> Department of Neuropsychology (A220)
> Stephanstraße 1a
> 04103 Leipzig
>
> Phone: +49 (0) 341 9940 2586
> Mail: kuhl at cbs.mpg.de
> Internet: http://www.cbs.mpg.de/staff/kuhl-12160
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://www.nitrc.org/pipermail/mrtrix-discussion/attachments/20150122/4e27aa99/attachment.html>