Hi Khoi,

Please, see below:

On 14 June 2016 at 03:57, Khôi Huỳnh Minh <[log in to unmask]> wrote:

Dear Anderson,

Thank you for you valuable reply. Since it is impossible to make the inference in this case, I should use GICA and have a visual inspection of the IC map instead.

I have three questions regard running the concat ICA.

I have 4 groups of subjects (training with strategy 1, training with strategy 2, training with strategy 3 and no training). There are two scan sessions for each subjects: before and after training. I want to look at the different between trainings method (group) and also the difference (if exist) after and before the training.

1. When performing the concat ICA, should I perform it on the whole data concatenated together such as:

subj 1 _ group 1 _ time 1

subj 2 _ group 1_ time 1

subj 1 _ group 2 _ time 1

subj 2 _ group 2 _ time 1

subj 1 _ group 3 _ time 1

subj 2 _ group 3_ time 1

subj 1 _ group 4 _ time 1

subj 2 _ group 4_ time 1

subj 1 _ group 1 _ time 2

subj 2 _ group 1_ time 2

subj 1 _ group 2 _ time 2

subj 2 _ group 2 _ time 2

subj 1 _ group 3 _ time 2

subj 2 _ group 3_ time 2

subj 1 _ group 4 _ time 2

subj 2 _ group 4_ time 2

Yes.

 

2. Given the set up of concat ICA above, could you give me an example how to design the design matrix and contrast for dual_regression. I usually make the design for 1 time multiple groups difference or 1 group multiple scan conditions differences but this is the multiple times multiple groups and things get complicated.

Please, have a look in the example given in this thread: https://www.jiscmail.ac.uk/cgi-bin/webadmin?A2=FSL;23e75b21.1601

It was for a similar case, except that it used 3 groups, whereas you have 4, so you'd make changes as needed. The design for that example (3 groups) can be accessed directly here: https://dl.dropboxusercontent.com/u/2785709/outbox/mailinglist/design_graham.ods

Note that there are two designs, such that you'd not run randomise as part of the dual_regression, but instead run it separately, by hand, twice after stage2, or use PALM (also after stage2)

 

3. Another problem is when performing dual_regression, is there an option to perform dual_regression on selected component. Say, I only want to perform it on IC components that are related to the executive networks not the default mode network. Since the number of subjects is large, given the number of IC is around 30-40 and the number of contrast I guess will be not smaller than 6, it will take a great deal of time to run randomise at dual_regression stage 3.

Yes, it's possible to run randomise separately (outside the dual_regression command) only on the IC selected.

All the best,

Anderson

 

Any suggestion would be very appreciated.

Sincerely,

Khoi

On Mon, Jun 13, 2016 at 2:10 PM, Anderson M. Winkler <[log in to unmask]> wrote:

Hi Khoi,

Please, see below:

On 12 June 2016 at 10:10, Khôi Huỳnh Minh <[log in to unmask]> wrote:

Hi Anderson,

Thank you for your prompt and valuable reply. Please consider my point below:

​I have recorded the time seed when subjects in my experiment tap their finger. I try to convolute this with HRF and find out that the result is somewhat correlate with one IC time series (correlation about 0.3-0.4). The spatial map of that IC show activation at motor cortex and frontal lobes. Can we conclude that the spatial map is the activation map for the "finger tapping" activities ? I doubt that by decompose fMRI signal into ICs, the IC time series do not have haemodynamic response characteristic.

Yes, that sounds the right interpretation. However, given it is a task-based experiment, and given that you know the stimulus onsets, perhaps a more powerful and successful approach would be to use the GLM, as opposed to ICA.

I have tried GLM but due to the complexity of the task (a maze escape game) and the fact that I only have the stimulus onset for finger tapping event (subject navigation)

​

 

​, it seems to be not very robust to perform a GLM on FEAT (because there are lots of activities happen in subjects' mind but I only have one EV). By using ICA, after unmixing, I hope that the model (stimulus convoluted with HRF) can be compare with IC time series given more robust result (since decomposing original BOLD response into IC time courses can help separate different activities in subjects's brain). My method is calculation correlation between the model and the IC time series to choose time courses that match with the event. I've also performed a GLMFIT (using Matlab) in addition to the correlation alone.

Using this approach, or correlating the IC timecourse with the stimulus timecourse will help to identify only those that are related to the known stimulus (i.e., the finger movements), and will not reveal other processes that were going on (such as related to the spatial navigation itself, memory, planning, etc). Perhaps these might even be among the ICs identified, but without a reference it won't be possible to identify them.

If you don't have that information, then it won't be possible to make that inference.

All the best,

Anderson

 

However, as we perform sICA for fMRI data, two or more time series can be together represent one brain activities (since we accept some correlation between IC series to get the maximum spatial independent). For each subject, I decide to choose some ICs that are highly correlated with the model, average them to get one spatial map. Then for group level, perform one sample t test randomise on a 4D data (each 3D data is spatial map from each subject).

Does 

​it seems a good approach​ or for each subject, I should not average spatial maps but just chose the one whose IC time series best match with the model ?

Best 

​regards, 

Khoi​

On Sun, Jun 12, 2016 at 3:14 PM, Anderson M. Winkler <[log in to unmask]> wrote:

Hi Khoi,

Please, see below:

On 12 June 2016 at 05:43, Khôi Huỳnh Minh <[log in to unmask]> wrote:

Dear Anderson,

Sorry for pull out this topic​

​ again but I have a question that related to what I mentioned before.

Is 

​the time series extracted from ICA have that characteristic of the haemodynamic response ? I means we all know that BOLD signal ​can be model by HRF and a set of stimulus and deconvolute BOLD signal with suitable stimulus can give us HRF. How about for time series from ICA result.

We would hope so, yes, after unmixing, the time courses still represent BOLD responses, that could, potentially, be subjected to deconvolution.

 

​I have recorded the time seed when subjects in my experiment tap their finger. I try to convolute this with HRF and find out that the result is somewhat correlate with one IC time series (correlation about 0.3-0.4). The spatial map of that IC show activation at motor cortex and frontal lobes. Can we conclude that the spatial map is the activation map for the "finger tapping" activities ? I doubt that by decompose fMRI signal into ICs, the IC time series do not have haemodynamic response characteristic.

Yes, that sounds the right interpretation. However, given it is a task-based experiment, and given that you know the stimulus onsets, perhaps a more powerful and successful approach would be to use the GLM, as opposed to ICA.

All the best,

Anderson

 

It would be appreciated if you can help me point out the problem as my conclusion make me feel somewhat fallacious.

Best regards,

Khoi

On Fri, May 13, 2016 at 3:27 PM, Anderson M. Winkler <[log in to unmask]> wrote:

Hi Khoi,

I understand that the null hypothesis is that there is no pattern across subjects, thus the expected average across subjects would be a map of all zeroes. This doesn't seem a very good hypothesis from the outset, as we may expect that commonalities among timecourses would be enough to render the maps similar among them. Still, such a test can be done with unthresholded maps, concatenated (in standard space) and tested in randomise with the option -1.

To make this a bit more objective, consider taking, for each component, the one that has the strongest correlation of timecourse with the stimulus function, even if such correlation is poor for some subjects.

That said, perhaps better is to simply use the GLM: since you have already a sequence of stimulus, these can be used as the regressors in a 1st level, and the common pattern can be found in a higher level. There will be no risk for circularity whatsoever, and no issues related to the scaling of components (z-stat, etc).

All the best,

Anderson

On 13 May 2016 at 04:34, SUBSCRIBE FSL Khoi Huynh <[log in to unmask]> wrote:

Dear FSL experts,

After running single ICA for all subjects, I find that each subject has 1 IC which its time series is highly correlated with my interest stimulus. The stimulus design is not the same for all subject hence I cannot use tensor ICA. I want to maintain as much as information possible so I dont want to use concat ICA (since concat ICA will run in MNI space instead of subject space).

Here is what I got after my single ICA run:
-Time series of IC 1 of subject 1 is correlated with the event subject 1 lost the game -> zstats_threshold map of IC1 - subject 1
-Time series of IC 7 of subject 2 is correlated with the event subject 2 lost the game -> zstats_threshold map of IC7 - subject 2
.... so on.

I registered all the zstats_threhold map to MNI space but then stuck at finding a way to find common pattern of them.
Hence, is there any way in FSL that I can find common activation from all of the zstats_threshold maps ? I am thinking of average all the map and threshold the result at a specific threshold but I feel like it is not the correct way.

It would be very appreciated if anyone can give me any advice.

Best regards,
Khoi