Well, I think you would be better off, as
far as ability to reconstruct multiple fibers and reducing your uncertainty of
orientation estimations, to use a single shell at the maximum bvalue. If
you are using the data for other purposes than FSL probabilistic tractography
that benefit from the lower shells, then that makes the picture more muddy. J
To test to see how you are doing, you can
compare your ability to reconstruct 2nd fibers to that of the sample
FSL dataset (at least I think FSL still distributes a 60 direction b=1000
sample dataset). You can generate a quantitative measure of 2nd
fiber reconstructed by determining the number of 2nd fibers with a
volume fraction above a certain threshold (e.g. 0.05, i.e. mean_f2samples
greater than 0.05) as a proportion of first fibers above that threshold.
For example, you could generate this ratio (#Voxels with mean_f2samples
> 0.05) / (#Voxels with mean_f1samples > 0.05) and compare between the
two datasets. If you just want a rough estimate, you can do this on the
files as they come out of bedpostx. If you want to be a bit more
accurate, you should make sure there aren’t any non-brain voxels or
voxels in which not all directions have data (e.g. what happens on the slice at
the edge of the FOV) after correction of subject motion and eddy currents
(usually done by eroding the brain mask to remove these regions from the
measurements you do with fslstats). For reference: on the 256 direction
dataset with b=1500 that was released for the Pittsburgh Competition I get 0.72
or 72% of white matter voxels have a second fiber above threshold. If I
recall correctly, this is similar to what one gets with a 60 direction b=1000
dataset, and the major differences are in number of 3rd fibers and
fiber uncertainty.
Uncertainty is a bit more difficult to
easily quantify with the released version of FSL, but I think the new version
will have the option to output the dispersion of the fiber estimates in
radians, though you may have to modify the bedpostx_postproc script to send
that flag to make_dyadic_vectors.
In general, at least with the released
version of FSL, one will do better as one uses more unique directions at a
single higher bvalue in reconstructing subsidiary fibers and reducing
uncertainty in fiber orientations. As to whether non-single shell schemes
can be exploited to do as well or better than single shell schemes is an
unsolved question!
Peace,
Matt.
From:
Sent: Wednesday, November 25, 2009
12:16 PM
To: [log in to unmask]
Subject: Re: [FSL] DTI with High b
value
Average
trace of bmatrix (sec/mm2) |
Number
of directions |
3.8 |
3 |
8.5 |
6 |
64.9 |
10 |
114.3 |
12 |
348 |
16 |
574.6 |
18 |
858 |
20 |
1198 |
22 |
This is the scheme we used for thalamic
segmentation, and I did not notice major abnormalities here with bedpostx,
however I might have to look more carefully. What would be a good test of this?
In this case, would it be advisable to drop
the intermediate shells (say b= 65-575?)? I thought that when we first
consulted Tim about this, he advised to maintain the intermediate shells to
gain SNR…. Stefano