Back to PLS HelpPosted on 07/02/13 12:48:54
Number of posts: 6
Dear plsgui experts,
I was using version 5.0 and now just shifted to v6.0 for a new behavioural PLS with quantitative images (quantitative T1 and quantitative T2 maps). I would appreciate advise for the following three issues:
1. Since the input maps are quantitative T1 and T2 maps, would it be better to go for 'PET' or 'Structural'?
2. During 'run PLS' I see there's a new cell to input '# of split' - could anyone advise what this is for and how is it different / is it dependent on the numbers of permutation and bootstrap? What would be the recommended parameter for this?
3. Assuming I should run Structural, should I deselect 'Segment' since I am not using segmented sMRI maps (i.e. GM, WM) here?
Thank you very much in advance!
regards.
Meng-Chuan
Replies:
I'm Online
Posted on 07/02/13 12:55:13
Number of posts: 229
Hi Meng-Chuan..
I would use the structural module - this allows "true" zeros, which are unlikely, but possible with your T1/T2 mapping methods. You need to have a brain mask that excludes outliers as well (i.e., voxels where the T1 or T2 fitting is bad, giving artifactual relaxation times). Yes, since these are not segmented images, do not use that option.
The "split" is referring to a new method using split-half reliability in the resampling step - I do not know if any documentation is available yet.
nancy
Posted on 07/02/13 13:04:34
Number of posts: 6
Hi Meng-Chuan..
I would use the structural module - this allows "true" zeros, which are unlikely, but possible with your T1/T2 mapping methods. You need to have a brain mask that excludes outliers as well (i.e., voxels where the T1 or T2 fitting is bad, giving artifactual relaxation times). Yes, since these are not segmented images, do not use that option.
The "split" is referring to a new method using split-half reliability in the resampling step - I do not know if any documentation is available yet.
nancy
Hi Nancy,
Thanks so much for the useful advice. Re split, is it the case that if it is left as default (0) and I simply set the permuation and bootstrap numbers, it would run as what was done in v5.0 and as suggested by the previous tutorial papers? Else would there be any suggestion re what could be tried on this new method?
regards,
Meng-Chuan
Posted on 07/02/13 15:26:18
Number of posts: 6
Hi Meng-Chuan..
I would use the structural module - this allows "true" zeros, which are unlikely, but possible with your T1/T2 mapping methods. You need to have a brain mask that excludes outliers as well (i.e., voxels where the T1 or T2 fitting is bad, giving artifactual relaxation times). Yes, since these are not segmented images, do not use that option.
The "split" is referring to a new method using split-half reliability in the resampling step - I do not know if any documentation is available yet.
nancy
Hi Nancy,
On another note, is it reasonable to have T1 maps and T2 map as two 'conditions' in the structural module Behavioural PLS? I recall that if the inputs are segmented maps we could have GM/WM as two or GM/WM/CSF as three conditions. Not sure if the same logic will apply to quantitative T1 and T2 maps (since they represent two indices of the MR property of the tissue in a particular voxel) ?
Thanks in advance!
regards,
Meng-Chuan
I'm Online
Posted on 07/02/13 15:41:22
Number of posts: 229
On another note, is it reasonable to have T1 maps and T2 map as two 'conditions' in the structural module Behavioural PLS?
absolutely - especially if you are running a behavPLS. (looking at the difference between the two relaxation times in a task analysis makes less sense.) Those are independent MR parameters, and while both can change in a complementary fashion, there is no reason they must - you will be able to see where the correlations with your behavioural measure are similar (or different) across the brain for those parameters.
I'm Online
Posted on 07/02/13 16:07:01
Number of posts: 291
To answer your first post regarding to the split:
As Nancy said, "The "split" is referring to a new method using split-half reliability in the resampling step", and there is documentation under: http://research.baycrest.org/pls/UserGuide.htm#_Toc349137832 (That section was from Natasha)
And you are right, left it as default (0) and you simply set the permuation and bootstrap numbers, it will run as what was done in v5.0 as suggested by the previous tutorial papers.
Posted on 07/05/13 08:05:00
Number of posts: 6
On another note, is it reasonable to have T1 maps and T2 map as two 'conditions' in the structural module Behavioural PLS?
absolutely - especially if you are running a behavPLS. (looking at the difference between the two relaxation times in a task analysis makes less sense.) Those are independent MR parameters, and while both can change in a complementary fashion, there is no reason they must - you will be able to see where the correlations with your behavioural measure are similar (or different) across the brain for those parameters.
Hi Nancy,
Thanks for the useful advice. I would like to grab this opportunity to further clarify the meaning of 'conditions' in structural PLS. My (heuristic) understanding is that a condition for a brain voxel in structural PLS represents a property of the voxel. So for example in the case of segmented sMRI maps, GM and WM can be different conditions because each voxel has a 'property' related to GM and WM, respectively (i.e., partial volume estimate, plus modulation so adding information about volume, etc.) Is this correct?
Then my next question is, do these properties need to be independent, that they can be put in as different conditions? From you reply I can see that since T1 and T2 are not depending on each other it is reasonable to put them in simultaneously. But if 'independence' is a requirement, then why would it be ok to have GM and WM (and CSF) since they might be rather interrelated? How about parameters in diffusion scans: I assume it will be ok to have FA and MD as 2 conditions, or even FA, RD and AD as 3?
The last related question is that I've encountered circumstance that if only putting GM as the sole condition there's no significant findings, neither does solely WM; but when both are included as two conditions a significant correlation pattern emerges. How do you think this could be understood?
Sorry if these are too general questions but I find it important to understand more about the reasoning behind.
Many thanks,
Meng-Chuan
I'm Online
Posted on 07/05/13 08:53:53
Number of posts: 229
these are good questions.. I'll try to answer...
PLS does not adjust for within block correlations, so having some correlation between "conditions" is not a problem.. The challenge with structural data is that in some brain regions, GM and WM values can be perfectly inversely correlated and in other locations that holds for GM/CSF. If two variables are perfectly correlated, then you would only need one to answer your question.
However, in our experience, it seems that once you've smoothed the data and/or modulated by the jacobian, those correlations are no longer perfect, so PLS is still able to identify reasonable correlations with external variables.
I would not use FA (tensor) maps in any voxel-wise analysis. This imaging modality is very sensitive to residual registration errors and edge effects - smoothing doesn't help much - see the papers by Derek Jones. On the other hand, you could enter TBSS data since the skeleton is well away from the edges of most fibre bundles (also keeping in mind that the registration scheme will also influence the 'accuracy' of the skeleton, and may still not give optimal mapping).
for your last question... that is not an unusual finding (and you've changed the covariance structure) but interpretation would depend on the location of the stable results and the pattern on the BehavLV. If the regions are indicating effects only at G/W interfaces (i.e., 50% probability of GM & 50% probability of WM), then you are likely identifying registration errors, and you should see equal and opposite correlations on the BehavScores correlation plots.
nancy
Login to reply to this topic.
