Regarding the thresholds of Rilling et al (2008), they were selected
qualitatively in several subjects, averaged, and then applied across the
whole group. The threshold was a percentage of the total number of
streamlines sent out, not the waytotal. The waytotal tends to be a much
smaller number than the total number of streamlines sent out, so you could
not just take the number we chose in that study and use it.
In a currently unpublished study of arcuate asymmetry, I tried a number of
thresholds (as a percentage of total number of samples) to see how it
changed the volumetric asymmetry of the pathway. I found that thresholds
between 5 x 10-7 to 5 x 10-6 gave volumetric asymmetries that correlated
with the asymmetries in the waytotals and a gold-standard measure particular
to these brains. The next lower and higher thresholds I tested, 1 x 10-7
and 1 x 10-5 produced results that were no longer consistent with these
other two measures. These threshold percentages all are for tractography
with 25000 samples (so multiply by 5 if you are using 5000 samples for your
tractography). The number we used in Rilling et al (2008), 3.8 x 10-5 / 5 =
6.2 x 10-6 (to convert to 25000 sample tractography), which is close to the
range that I found above. I did not have success thresholding by a
percentage of the waytotals (volumetric asymmetry did not correlate with the
other measures), however perhaps this was due to the bug that has been
discussed elsewhere on the list.
Peace,
Matt.
-----Original Message-----
From: FSL - FMRIB's Software Library [mailto:[log in to unmask]] On Behalf
Of Cherif Sahyoun
Sent: Friday, November 28, 2008 1:45 PM
To: [log in to unmask]
Subject: Re: [FSL] AW: [FSL] waytotal thresholding
Hi John,
To add to Andreas' reply, a simple way to think of it is thresholding
allows you to get rid of voxels with a low probability of lying on the
tract (so as Andreas said, the important thing is to be consistent).
On the other hand, dividing by a given number will normalize the
datasets, so that values are comparable across fdt_paths/subjects.
Imagine in one subject your seed is twice the size of that of a second
subject, when you get many more tracts, it doesn't mean that the
pathway is more probable (not yet!), unless you adjust for say the ROI
size...
Hope that helps,
Cherif
----------------------------------------------------------------------------
--------------
Cherif P. Sahyoun HST-MEMP
Developmental Neuroimaging of Cognitive Functions
C: 617 688 8048
H: 617 424 6956
[log in to unmask]
"Live as if this were your last day. Learn as if you'll live forever"
Ghandi
----------------------------------------------------------------------------
---------------
On Thu, Nov 27, 2008 at 3:07 PM, Andreas Bartsch
<[log in to unmask]> wrote:
> Hi John,
>
> I think it does not really matter (or no one can tell;) but you should be
consistent (i.e. use a predefined percentage). For clinical purposes, I
often felt comfortable with 0.1% if (!) I wanted to threshold.
> Similarily, normalizing fdt_paths to the waytotal does not alter the
spatial extent of the original output. It just "scales" the values (to
interpretable probabilities if you know that the tract you were after
exists). Thresholding will change the spatial extent of the tract. So if you
threshold at 0.1%. for example, you will exclude those voxels that have only
a chance of 1 in a thousand that they belong to the tract of interest (given
you have good reason that the tract must be there).
> Hope that helps,
> Cheers-
> Andreas
>
> ________________________________
>
> Von: FSL - FMRIB's Software Library im Auftrag von John D. Griffiths
> Gesendet: Do 27.11.2008 16:17
> An: [log in to unmask]
> Betreff: [FSL] waytotal thresholding
>
>
>
> Hi Tim + Matt +all.
> With regard to this notion of thresholding as a
> consistent % of the waytotal number: Roughly what kind of numbers have
> people been using to do this? I know you need to choose one that looks
best,
> etc., but I'm wary of being too 'liberal'...
>
> In Matt & Tim's nature neuroscience paper (Rilling et al. 2008), they
> threshold to include only those voxels that received 'at least 0.000038%
of
> the total streamlines sent out from the ROI masks'. Now, given that the
> analogy has been made several times on this mailing list between
> thresholding as a proportion of the number of seed voxels and thresholding
> as a proportion of the number of streamlines passing between two seeds
(i.e.
> the waytotal number) - would it be ok to threshold my results at values as
> low as 0.000038% of the waytotal?
>
> Could anyone point me to some papers that have used this thresholding as a
> proprtion of the waytotal approach?
>
> Finally (and sorry if this is a really dumb question...) - One suggestion
on
> this list has been to normalize probtrackx results by diving by the
waytotal
> number (which I read as 'fdt_paths -div (waytotal)'), whilst others have
> been to threshold as a consistent % of the waytotal number (which I read
as
> 'only include intensities greater than or equal to [(waytotal /
100)*(e.g.)
> 10]'). Are there any important differences between these?
>
>
>
> Thanks,
> John
>
|