Hi,
please let me add a comment first:
Matt suggested that thresholding fdt_paths normalized to the waytotal is good to get consistent trackings ACROSS subjects whereas thresholding fdt_paths normalized to the total number of samples send out is good to detect tract asymmetries WITHIN subject. Although I can see the point he is trying to make, it is - in my opinion and experience - not entirely true.
Thresholding a tract by its waytotal percentages enforces a more uniform "trackability". The original rationale why Tim and I came up with this is that you may have other evidence that the tract exists, e.g. preserved motor function indicating some (residual) integrity of the pyramidal tract in brain-lesioned patients. Imagine a case where the pyramidal tract in one hemisphere passes through the perifocal edema of a tumor. The edema massively increases the diffusion uncertainty and much fewer samples reach the target from the seed. However, the patient has no motor deficit and the pyramidal tract is only embedded in the edema but not infiltrated or distructed. If you now compare your volumes based on thresholding the pyramidal tract by percentages of total number of samples send out you will end up with great asymmetry which is essentially a false-positive detection. This may be eliminated or is, at least, attenuated when you thresholding by waytotal percentages. Heidi and Tim have a book coming out where we present a chapter with clinical data illustrating and further explaining this.
So my point is that thresholding tracts by their waytotal percentages attempts to enforce a more uniform "trackability" ACROSS OR WITHIN subjects based on other prior evidence that the tract must be there. Thresholding fdt_paths normalized to the total number of samples send out may, on the other hand, overestimate tract asymmetries within subjects if there is an interhemispheric "trackability" difference. Ok, brain lesions are obvious. But you need to ask yourself if you are sure that "trackability" is the same between the hemispheres of the same subjects. In epilepsy patients, for example, there may be less obvious microstructural differences. So the waytotal is good if you have easily accessible and measureable evidence that a tract exists (which is the case for the pyramidal tract or the optic radiatio, for example). It would be much harder for the arcuate fascicle, esp. in the right hemisphere. For many tracts, we have to keep in mind that there is no good and easy way to tell (aside from our tracking!) if they do in fact exist.
Ok, now to your questions:
>ROIs symmtrical between both sides
When the anatomy is asymmetric, there is no point to make them symmetric in shape. However, you may be able enforce (roughly) the same size. We did so for the data presented in the upcoming book chapter mentioned above.
>So if I don't reject any tracts, the waytotal should be the same as total number of samples?
Yes.
>but will it confuse our readers if we use different % of threshold in one article
It certainly may. Preferably, you may have to explain based on another sample you've studied why you have decided for a particulat threshold for a given tract.
Hope that helps,
cheers-
Andreas
________________________________
Von: FSL - FMRIB's Software Library [[log in to unmask]] im Auftrag von cathyliu [[log in to unmask]]
Gesendet: Montag, 30. März 2009 11:21
An: [log in to unmask]
Betreff: Re: [FSL] asymmetires
Thank Matt for sharing your experiences!
Some other things to think about regarding this (to add what I said below): Bigger tracts will often have bigger masks, and if you are really interested in structural asymmetries, you want to be sure that you get the entire pathway for each of your subjects. If you leave something out, you could either create a false asymmetry or miss a true one. Making sure you get the whole pathway (but only the whole pathway), is more important than keeping the size of your masks exactly the same in my opinion. At the same time, if you are using the size of the ROI as a part of your thresholding or normalization scheme, you need to be sure that you don’t have “extra” voxels in your ROIs, i.e. voxels that don’t track the pathway, or don’t track anything at all. Example: You have ROI A of 100 voxels in which 50 voxels track a pathway of interest. You have ROI B of 50 voxels in which all 50 voxels track the pathway of interest. If you are generating thresholds based on the size of your ROIs, the pathway from ROI A will be more stringently thresholded than the pathway from ROI B.
<<<< We put relatively bigger masks in order to get the entire pathway, and meanwhile we try to make the ROIs symmtrical between both sides. But as you know the brain structures are asymmetrica, so it is impossible to make the ROIs both sides exactly the same sizes. By the way, if we use waytotal-based thresholding, could it avoid the effect of different size of the ROIs?
I personally tried both waytotal-based thresholding and thresholding based on total number of samples (# voxels in the ROI X # of samples sent out per voxel) on a study of pathway asymmetry. I found significantly better correlation between a gold standard measure of functional lateralization and tract volume asymmetries generated from thresholding based on total number of samples than on tract volume asymmetries generated from waytotal-based thresholding. I observed that waytotal-based thresholding tended to make pathway volumes more consistent across subjects and hemispheres, thus making it harder to see interhemispheric differences.
<<<< I read the information from the mail list again. If I understand well: waytotal= # voxels in the ROI (fslstats -V) X # of samples sent out per voxel(5000 by default) if no tract is rejected. So if I don't reject any tracts, the waytotal should be the same as total number of samples?
Regarding normalization vs thresholding, you will have to do thresholding regardless of whether you normalize. If you normalize, you can just use the same number as your threshold for all subjects, but you lose information about the absolute pathway intensities. If you don’t normalize, you will need to calculate the threshold separately for each pathway (i.e. decide on some proportion of the waytotal or total number of samples tracked, then for each pathway, multiply the proportion X the waytotal or the total number of samples tracked to get the threshold value for that pathway). I think this is more of a question of personal taste, rather than right or wrong, as both methods are equally valid for interhemispheric comparisons. If you want to compare across subjects, however, it is better to normalize (and to normalize using wayotals), as this can account for global differences between scans (e.g. more motion).
<<<< If we are interested in series of tracts, some of them are across subjects, but some of them are not, should we keep the same way of thresholding? As you mentioned in last mail, we could choose different % of threshold for different tracts, but will it confuse our readers if we use different % of threshold in one article?
Best Regards,
Yan
________________________________
From: FSL - FMRIB's Software Library [mailto:[log in to unmask]] On Behalf Of cathyliu
Sent: Wednesday, March 25, 2009 9:28 AM
To: [log in to unmask]
Subject: [FSL] asymmetires
Hi,
We want to compare the structural asymmetries between the right and left hemisphere by probabilistic tractography. We still have some concerns: if the size of the masks will have an effect on the results? For example, a big mask will have more tracts, and in fact we couldn't make the masks exactly the same on the both sides. The second, for comparing the results, if it is necessary to do the normalization and what is the function of normalization? In one of your emails, it mentioned that:
My understanding is that dividing by the waytotal is the way to normalize
across subjects as best as you can. Then you could threshold at some
fractional proportion of the waytotal and binarize. This should give you
the most similar spatial distributions across subjects. However, I think
that this method is less useful when you are making comparisons within a
brain, for example between the same pathway in the two hemispheres, when you
are interested in pathway asymmetries for example.
So what is the best method, if we are interested in pathway asymmetries?
Best Regards,
Yan
________________________________
把MSN装进手机,更多聊天乐趣等你挖掘! 立刻下载!<http://mobile.msn.com.cn/>
________________________________
把MSN装进手机,更多聊天乐趣等你挖掘! 立刻下载!<http://mobile.msn.com.cn/>
|