Hi Derek,
I did option number 2 to get around the
thresholding problem (it is very hard to figure out exactly how to define a
threshold in a principled way, see numerous discussions on the list about this).
I did it on whole ROI results, though, which is not what he is asking
for. I did this on symmetrically tracked results from two ROIs, which
should only bias measurements towards tracts that are between the ROIs. If
you used single ROIs or asymmetric tracking (seed à waypoint) you would
probably bias your results towards the seed ROI. As compared to
thresholding the paths at some value, I found that I got higher FA values from
this method, likely because high probability voxels tend to be in core white matter.
Unfortunately, this was for a grant submission so I don’t have a
citation for this method.
Peace,
Matt.
From: FSL - FMRIB's Software Library
[mailto:[log in to unmask]] On Behalf Of Derek
Jones
Sent: Thursday, October 21, 2010
10:37 AM
To: [log in to unmask]
Subject: Re: [FSL] Probtrackx:
Vector Output?
I'd also like to
ask - for those that have employed such an approach:
1) Do you choose
a statistical threshold on the ProbTrackX result - and then take the average
metric from all remaining voxels in the tract (i.e. 'significant'
tracts). If so - what threshold do you use?
2) An alternative
approach would be to weight the metric by the 'probability' of the voxel being
part of the tract, such that those voxels that are intersected most
frequently will have a proportionately larger contribution to the average.
...and does this
bias the measurements to the FA / RD / ADC of the tract that is closer to the
seedpoint?
Thanks in advance
to anyone that replies
I'm wondering if
any of you have found a way to export the results of probtrackx as paths,
rather than as images. I'm interested in looking at structural connectivity in
a voxelwise manner; e.g. given two voxels, looking not only at the relative
probability of a connection, but also at the average microstructural measures
(FA, RD, ADC) of the paths found. This should be possible (since the paths must
originally be calculated in these terms) and given the popularity of this
approach, I figured someone has probably done it already with FDT.
Just using the
fdt_paths output as a mask won't help to answer the question because I want to
examine compare voxels within seed masks, rather than across them (and running
each voxel individually as a seed would be very inefficient).