Hi Jian,

Please, see below:

On 11 May 2016 at 15:48, Jian Zhang <[log in to unmask]> wrote:
Hi Anderson,

Thank you for the reply. If we want to use TFCE for inference, since VBM and resting state use -T -tfce1D option, TBSS uses -T -tfce2D option,

Something isn't right above. Volume-based data (3D) should go with -T alone, without -tfce1D (that would be for 1D data, such as time-series, e.g., time-resolved NPC on M/EEG data).

 
can we run the command like this:

palm -i all_FA_skeletonised.nii -T -tfce1D -i GM_mod_merg_s3.nii -T -tfce1D -i dr_stage2_ic0000.nii -T -tfce2D -i dr_stage2_ic0001.nii -T -tfce2D -i dr_stage2_ic0002.nii -T -tfce2D -d design.mat -t design.con -n 5000 -corrcon -corrmod -save1-p

This won't work I'm afraid. The -T counts only once for all modalities, so these multiple -T will have no extra effect. The -tfce1D and -tfce2D also count only once (the last entered), such that in this case, it will use settings for 2D, as it's the last in the command line.

Another issue is conceptual: TFCE isn't a pivotal statistic as it depends on the space surrounding a given voxel (i.e., its neighbours). The neighbours of a given voxel in the skeleton are a different set than the neighbours of the same voxel in a non-skeletonised image (e.g., VBM, fMRI), such that it isn't possible to correct here (in the FWE sense) across all these modalities if TFCE is used, but it is possible to correct in one run the skeletonised data only (FA, MD, etc), and in another run the non-skeletonised only.

Alternatively, it's possible to correct using FWE across all modalities if TFCE is dropped. So, any of these below will work:

1) Skeletonised data only, with TFCE 2D. With just 1 modality, -corrmod can be removed (but add back if other skeletonised modalities are included, e.g., MD, etc.):

palm -i all_FA_skeletonised.nii -T -tfce2D -d design.mat -t design.con -n 5000 -corrcon -save1-p

2) Non-skeletonised only, with TFCE:

palm -i GM_mod_merg_s3.nii -i dr_stage2_ic0000.nii -i dr_stage2_ic0001.nii -i dr_stage2_ic0002.nii -T -d design.mat -t design.con -n 5000 -corrcon -corrmod -save1-p

3) Everything, without TFCE:

palm -i all_FA_skeletonised.nii -i GM_mod_merg_s3.nii -i dr_stage2_ic0000.nii -i dr_stage2_ic0001.nii -i dr_stage2_ic0002.nii -d design.mat -t design.con -n 5000 -corrcon -corrmod -save1-p


Or do we have to register every modality to the same space ( standard space, e.g.) and then using the command of NPC:

palm -i all_FA_skeletonised_standard.nii -i GM_mod_merg_s3_standard.nii -i dr_stage2_ic0000_standard.nii -i dr_stage2_ic0001_standard.nii -i dr_stage2_ic0002_standard.nii -d design.mat -t design.con -n 5000 -npc -Tnpc -save1-p

NPC is related to correction, yet it's a different thing. For NPC, all images need to be in the same space, and PALM will produce an intersection mask across all the modalities, and combine only those (and also run the partial tests only for those). Here, with modalities that focus on GM (VBM) and WM (FA), the overlap is minimal, and there won't be much left to analyse.

 

For corrected across, "Each modality will be tested separately", so is it just for estimation of one modality after regressoring out variables of other modalities?

The correction uses each modality tested separately, but crucially, using permutations that are done synchronously across all, such that the correction considers their non-independence. There isn't leak of information from one modality into another during the regression stage.

 
If we want to estimate the interaction between modalities, should we transfer to joint inference (NPC or MANOVA/MANCOVA)? Just like FSLNets, explore the correlation between 2 RSN, using PALM to search the correlation between 2 modalities, for example, change of modality A causing positive/negative alteration of modality B?


Yes, this is possible through the use of voxelwise EVs. This example, on whether modality A is associated/correlated with modality B, can be done using A as input (dependent variable, -i), and B as a voxelwise EV in the design (or vice-versa). It can be done in either PALM or randomise.

 
And estimating the contribution of modality A and modality C to the alteration of modality B?

Yes, also possible. For this, use B as input, and prepare two designs, one having A as voxelwise EV, another with C as voxelwise EV, then run these two designs in PALM with -corrcon, and at the end, use fslmaths to find the largest corrected p-value across both, which will indicate a conjunction (IUT in the paper).

Alternatively, for a joint test (UIT), use -npccon, that will do NPC over the contrasts for these two designs.

 

Moreover, can we use palm to find similar or same changed brain area during modalities, and the link between them:
Modality A shows change in area A (near area B)/ area B, modality B shows change in area B, estimating whether the similar/same change (in spite of the indices for each modality are different) own some relationship.

Yes, this seems a case for NPC over modalities (-npcmod). Can be done in PALM, and works regardless of modalities being different (e.g., VBM and resting FMRI). However, they must be in the same space and have the same resolution.

Hope this helps.

All the best,

Anderson

 

Is it possible?

Kind regards,
Jian