[Mrtrix-discussion] best parameter with a long-time scan?

[Donald Tournier]

Hi Romain,

In terms of the number of directions, I don't think it'll be worth
increasing the number beyond around 60 - I had an ISMRM abstract about this
a while back (which I'll hopefully get published at some point soon)
showing that you really can't see much in the SH terms above lmax=8.
There's just a suggestion of an lmax=10 term at b>=3000, but it'll be
drowned in the noise for any realistic protocol. That mean you'll need at
least 45 directions (for lmax=8), and I'd recommend going beyond that
minimum number to make sure the problem is comfortably over-determined.
This is particularly important if you're going to perform motion-correction
(which you should given the length of your scan) with rotation of the
gradient directions, a process that will tend to reduce the uniformity of
your gradient directions. That said, you'll probably find it
near-impossible to perform good motion correction - more on that later...
You might want to push the number of directions beyond 66 if you really
want to get the lmax=10 terms (maybe ~80 to make sure it's
over-determined), but I'm not sure it'll necessarily make much of a
difference. Otherwise, I'd recommend you repeat those directions, which at
least gives you the potential to do something with the redundancy...

A bigger problem you'll face is the consequence of the low SNR you'll have
given your small desired voxel size and higher b-value. While I'd consider
that b-value to be optimal in terms of contrast-to-noise ratio, it does
cause other problems when the SNR is low: Rician bias, and poor performance
on just about any motion correction algorithm. The problem with the Rician
bias is that even with a lot of averaging, the noise floor on magnitude
data will push up the low signal intensities, and this tends to reduce the
contrast between low SNR amplitudes. If you have a small voxel size (low
SNR in each DW image), then the noise level will probably be of the same
order as the DW signal, and this will tend to flatten out the angular
contrast. The other problem is that this non-zero background signal will
tend to introduce noisy peaks in the FODs, since the CSD process will
interpret them as signal and try to fit lobes to match the noise - you'll
probably see lots of noisy peaks in the ventricles for example. There have
been a few methods proposed to rectify the Rician bias, and you may find
that they can help, but I have no experience with this.

Motion correction is another problem in high b-value HARDI, because of the
low SNR on the one hand, and because the contrast between the different
images is further exaggerated, with makes life very difficult for the usual
image matching metrics (mutual information, typically). In my experience,
most methods currently available introduce more artefacts into the
reconstruction than they solve. We had an abstract on a mask-based method
to perform robust motion correction on high b-value data at this year's
ISMRM (Dave Raffelt) and Jesper Andersson also had a method based on
Gaussian processes. You might find that one or both of these approaches
could be used, but as far as I know neither is currently available. We're
planning on making the mask-based method available in a future version of
MRtrix, but it's not ready yet...

One much easier option for motion correction when you're confident your
subject won't move suddenly, is to intersperse b=0 images every ~10 DW
images, so you can estimate the motion based on these b=0 images and
interpolate the motion parameters to apply them to the DWIs. This is the
approach I'd originally used for my 2004 spherical deconvolution paper,
which also was a ~1h acquisition, and that worked pretty well. Not robust
enough for patients though, they'll tend to move in jerks, which introduces
higher frequencies in the motion than you can capture with the b=0 images,
and also tends to corrupt the signal due to intra-volume mis-registration
and signal dropout.

Hope that helps.
Cheers,

Donald.


On 14 August 2012 22:19, romain quentin <rom.quentin at gmail.com> wrote:

> Dear experts,
> I'm using CSD from mrtrix with a diffusion sequence with b = 1500, 32
> channels coil, 60 directions and isotropic 1.7mm voxels. This sequence
> takes about 15 minutes.
> I want to acquire a better diffusion sequence with more scanning time
> (around 1 hour). I think to use b value = 3000 but i'm not sure about the
> voxel size and numbers of directions.
> Do you think it's better to do 4 x 60 directions, 2 x 120 or around 200
> directions of diffusions?
> And I know that reduces the voxel size is risky for the SNR but do you
> think that I can keep a 1.7 isotropic regard to the time scanning (1 hour)?
> Thank you very much for your help.
> Best regards,
>
> Romain QUENTIN
> rom.quentin at gmail.com
> 01 42 16 00 67
> 06 80 33 15 74
>
> _______________________________________________
> Mrtrix-discussion mailing list
> Mrtrix-discussion at www.nitrc.org
> http://www.nitrc.org/mailman/listinfo/mrtrix-discussion
>



-- 
Jacques-Donald Tournier (PhD)
Brain Research Institute, Melbourne, Australia
Tel: +61 (0)3 9035 7033
-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://www.nitrc.org/pipermail/mrtrix-discussion/attachments/20120815/3f302040/attachment.html