Hi,
my figure so I give it a shot:
In that figure the waytotals were NOT used directly. Instead I used the ratio of the waytotals from the tracking in one hemisphere diveded by the other (see the legend). Because masks were equally sized in both hemispheres (p. 433 in the book), there was no need to 'normalize' further. So this would be an easy way to to it. Still I agree with Saad: anatomicofunctional correspondence should rule over mask sizes - size does not matter!
Also note that I've excluded two outliers. Overall, in my experience the waytotals can easily differ by a factor of 5 between hemispheres of the same subject and that must not have any biological meaning even if you have high-quality scans. Still, as the figures shows normalizing to the contralateral hemisphere may be useful.
Note that my measure was neurological, behavioural scores may show weaker or no correlations.
Hth-
Andreas
________________________________
Von: FSL - FMRIB's Software Library [[log in to unmask]] im Auftrag von hd x [[log in to unmask]]
Gesendet: Donnerstag, 10. Februar 2011 11:17
An: [log in to unmask]
Betreff: Re: [FSL] Normalizing waytotal itself?
Dear Saad,
Thanks a lot for your detailed reply. However, I didn't want to use the normalized fdt_faths to represent the connectivity strength, but use the waytotal directly as suggested by "Tractography for surgical targeting in Heidi's & Tim's book. Figure 19.12". This is also suggested by Matt in one of his replies to this maillinglist: "What is it that you are trying to show? That subjects with a stronger/weaker pathway have some difference in behavioral scores? Honestly, I might correlate the waytotal values themselves (if you are using at least one waypoint mask) with the behavioral score."
Thus my question was if I shall normalize Waytotal itself by the sum of the seed and target sizes. But not how to normalizde fdt_path by waytotal. I'm sorry I didn't made this very clear in my first email in the background of so many discussion about the latter in this maillinglist.
I really hope I can get some comments on that since I didn't see much discussion about it yet.
Best,
Huadong
On 9 February 2011 12:18, Saad Jbabdi <[log in to unmask]<mailto:[log in to unmask]>> wrote:
Dear Huadong,
1) Waytotal normalisation
When you run probtrackx, you are effectively building up a distribution using sampling, in the same way you may build a histogram by counting samples that fall into categories (bins).
So for example, the fdt_paths file gives you the histogram for the spatial distribution of connections, and waytotal counts the number of samples that satisfy certain conditions (waypoint/exclusion criteria). When you divide the fdt_paths results by waytotal, you are acknowledging the fact that fewer samples have been used to build up the spatial histogram than the maximum possible number (i.e. #seed voxelsX#samples per voxel). You need to do that in order for your histograms to be comparable across subjects/hemispheres. In the same way, if you have a histogram of counts where you've used 100 samples, you cannot compare the values to a histogram that used only 50 samples, you can compare the odds but not the raw values.
2) Mask size and other concerns
Although many people think that the mask size bias is a big worry, I personally think there are other more fundamental issues.
When you calculate the probability of a connection from a seed to a target, then if the target is appropriately matched across subjects/hemispheres (e.g. it is the same anatomo-functional entity), then you are comparing like with like, so you shouldn't worry about mask size. However, it is true that the probabilities may be higher for bigger target masks, which is a problem that comes from the very nature of the probabilities that you calculate.
In probtrackx, you are _not_ calculating tract "strength" (whatever that is), but instead you are calculating the level of confidence that you have on the trajectory from the seed to the target (of course through the high diffusion directions, goes without saying!). So you would expect your confidence to be higher if, for example, your target area is bigger (other things may also increase or decrease your confidence). If that is not satisfactory when it comes to regressing against behavioural data, it is not because of the size bias, but because of the quantity that you are calculating, ie your confidence on the connection.
So even if you manage to somehow take into account the size of the mask in the probability (eg if you had a model for the conditional probability on mask size), you are still faced with the problem that a higher confidence on a connection does not necessarily imply a stronger functional connection between two areas. You may need to consider other quantities (e.g. microstructural or functional models), that you may calculate using the tract distribution, and disambiguate low confidence on tract from low functional connectivity etc.
Cheers,
Saad.
On 9 Feb 2011, at 10:44, hd x wrote:
Dear list,
From previous discussions, I got the impression that it's good to use waytotal as the connectivity index when you want to correlate pathway strength with individual behavoural performances and when you are using at least one waypoint mask. (See http://mail.google.com/mail/?shva=1#search/label%3Afsl+interpretation+waytotal/125995c081e68cd7
http://mail.google.com/mail/?shva=1#search/label%3Afsl+threshold+pathway/12a2b2c25aec6d52)
I'm now working on a study which investigates the correlation between pathway laterality and individual behavioural performances. I have a seed mask, a waypoint mask and a target mask (which is the same brain area as the waypoint mask covers) in each of the hemispheres. The right masks are homologous areas to the left masks. Then with the left masks I traced my left pathway and the right masks the right pathway. I used the waytotal of the pathway to represent the pathway strength and calculated the pathway laterality accordingly by taking the absolute difference between the left and right waytotal.
However, I got another concern. The masks I used for the left hemisphere are not of the same size of the right ones, although all subjects used the same set of left and right masks. In this case, I'm wondering if I should normalize the waytotal based on the mask size before I calculate the lateralization index. If I should, what would be the right way to do it? Would it be good to divide the waytotal by the sum of the size of seed and target masks?
Your inputs will be highly appreciated.
Best,
Huadong
--
Saad Jbabdi
University of Oxford, FMRIB Centre
JR Hospital, Headington, OX3 9DU, UK
(+44)1865-222466 (fax 717)
www.fmrib.ox.ac.uk/~saad<http://www.fmrib.ox.ac.uk/~saad>
|