Hi Martin,
>Different fibers may have different "trackability" and using the same
threshold may thus yield over or under estimation of the track volume and
show/hide intrahemispheric asymmetry.
absolutely! The safest is not to threshold at all but just to use the waytotal to convert the fdt_paths into probabilities given that you have evidence that the tract exists. Mathematically, the probablilty you get is interpretable (it reflects the probability that a given voxel is on the tract given that you know that the tract exists). However, physiologically it may be much tougher to interpret and depending on what you are after and how you have tracked it you may want to threshold to exclude less plausible parts.
Peak, mean or probabilities for specific voxels extracted from fdt_paths normalized to the waytotal are often easier to interpret (e.g. you can show that the probability to assign a given voxel to the pyramidal tract must not be reduced by perifocal edema if you have prior knowledge that the tract is still there). But in a way that is just a mathematical trick and what you really want to know is how to sensibly threshold.
>comments and thought on this approach and possibly how
to go further with it.
Very good thoughts and I tried about the same but so far wasn't really sucessful, unfortunately. Sensible thresholds seem to vary and we don't really know what is relevant in the distribution. At least, I haven't been able to come up with good fittings.
I think the previous discussions with Matt nicely summarize what the waytotal normalization may be good for but so far I can't advice on how get 'good' thresholds. However, it is certainly rather in the one-tenth of percent than the percent magnitude!
Keep me posted!
Cheers-
Andreas
________________________________________
Von: FSL - FMRIB's Software Library [[log in to unmask]] im Auftrag von Martin Kavec [[log in to unmask]]
Gesendet: Montag, 6. Juli 2009 11:30
An: [log in to unmask]
Betreff: [FSL] Idea about thresholding probtrack
Hi,
with respect to our tractography research project I had a discussion with a
colleague about thresholding the probtrack results normalized by waytotal. We
are not particularly conviced by using arbitrary fixed threshold, i.e. 5%,
10%, or even as low as 2%, as has been discussed in this mailing list as
well. Different fibers may have different "trackability" and using the same
threshold may thus yield over or under estimation of the track volume and
show/hide intrahemispheric asymmetry.
Instead, what we thought could be used in place of the fixed threshold is
following:
1. for each probtrack output, create a histogram of the track probability
(probtrack normalized by waytotal), which could represent a "track
probability distribution function". Even the full set of values could be used
here, without binning.
2. fit the probability distribution function to a known distribution
functions, i.e. exponential, gamma, normal
3. use a descriptive statistical parameter (percentile, STD, ...) of a given
distribution as a threshold.
The problem with this approach is in 2. that the "track probability
distribution looks" very much like double exponential, with steep initial
decay and very long tail and I haven't been able to test it successfully
against either exponential nor normal distributions using K-S test.
I would appreciate the comments and thought on this approach and possibly how
to go further with it.
Thanks,
Martin
|