open-discussion > Welcome to Open-Discussion
Showing 1-8 of 8 posts
Display:
Results per page:
Jun 1, 2015  12:06 AM | Tawfik Moher Alsady
Welcome to Open-Discussion
Welcome to Open-Discussion
Jun 8, 2016  04:06 AM | Sara De Simoni - Imperial College London
RE: Welcome to Open-Discussion
Dear mICA toolbox developers,

This toolbox is a great idea and I haven't had many problems either with the installation or with initial use. However I definitely have some questions (perhaps this is also my inexperience!) and it would be extremely helpful to have some documentation/manual (including how to run the steps on the command line) to help navigate the toolbox.

I have a few questions if that's ok:

(1) Is it always necessary to input the files to be processed manually? I have been running the mICA steps through the command line to avoid doing this but I was wondering if there was a way to include the text file in the GUI as well.

(2) I am running group mICA on the HCP rest data - I am using a priori defined binarised striatal mask. However whether I run this on 10 or 100 subjects I get the following melodic error: No convergence after 500 steps. This only seems to be a problem for certain dimensionalities (I am running 2-20) and I initially missed this error as all the output was created.

Is this an error that you came across? Do you think it is because of the data (it looks fine to me) or some of the input parameters I am using (although I am essentially using just the defaults).

Thanks very much for your help, I would greatly appreciate it as I think this toolbox offers some really great stuff and would love to be able to use it fully!

Thanks,
Sara
Jun 8, 2016  08:06 AM | Sara De Simoni - Imperial College London
RE: Welcome to Open-Discussion
Specifically the error I am getting is this when running the group ICA option:

seq: invalid floating point argument: /group/HCP/.../dim4/melodic_IC.nii.gz
Try 'seq --help' for more information.
No convergence after 500 steps
Jun 9, 2016  06:06 AM | Florian Beissner
RE: Welcome to Open-Discussion
Dear Sara,

thank you very much for your feedback. It is greatly appreciated.
The current version the toolbox (1.12) does not support importing design files. We will try our best to add this functionality in a future release.

The second problem that you are describing (no convergence) happens every now and then and is hard to predict. When running a reproducibility analysis for a number of different dimensions, we have an output called no_convergence_error.png indicating, how many of the ICA repetitions for a certain dimensionality did not converge. In general nonconvergence is a hint that the chosen dimensionality does not describe the data well enough. I would exclude those dimensionalities for further analyses.

Hope that helps!

Florian
Jun 9, 2016  06:06 AM | Sara De Simoni - Imperial College London
RE: Welcome to Open-Discussion
Dear Florian,

Thanks very much for your answer! Yes when I run the group ICA, in many instances it seems to fail on dimensionality 4 - so I will exclude this from further analysis. However if I run melodic (through standard FSL commands) I do not get this error. Do you know why this may be?

When I run the reproducibility analysis I appear to get an error associated with dimensionality 11 (the corr_dim11.txt file lacks some information) - which I think is reflected in the following error? (this is from the error files of the split half step 3, whereas step 2 in some instances gives me the 'non-convergence' error).

Traceback (most recent call last):
File "/share/apps/mICA/mICA_Toolbox/py/ic_corr.py", line 131, in
value = cor_mat[row][column]
IndexError: list index out of range

My questions are:
(1) Can I still use this analysis but just ignore the dimensionalities that appear to cause problems?
(2) Why does the groupICA cause problems with dimension 4 but the reproducibility analysis causes a problem with dimension 11? Is this unpredictable as you said?
(2) I unfortunately couldn't find the no_convergence_error.png file - could you give me an indication of where this may be saved?
(3) This may be a silly question but am I correct in assuming that to create a graph such as Figure 2(A) in your paper, I would just need to average the numbers in the corr_.txt files for each dimension?

Thanks very much for your help and sorry for the multiple questions!
Sara
Jun 9, 2016  10:06 PM | Florian Beissner
RE: Welcome to Open-Discussion
Dear Sara,
Thanks very much for your answer! Yes when I run the group ICA, in many instances it seems to fail on dimensionality 4 - so I will exclude this from further analysis. However if I run melodic (through standard FSL commands) I do not get this error. Do you know why this may be?
When you say you run melodic, does that mean you run melodic with the same mask, same resolution and smoothing you use in the toolbox? Or do you mean an unmasked ICA? In the latter case it would not be surprising to get completely different results because of different intrinsic dimensionalities of the different brain regions. Unfortunately, we had to remove that part of the results from our toolbox paper prior to publication.
When I run the reproducibility analysis I appear to get an error associated with dimensionality 11 (the corr_dim11.txt file lacks some information) - which I think is reflected in the following error? (this is from the error files of the split half step 3, whereas step 2 in some instances gives me the 'non-convergence' error).

Traceback (most recent call last):
File "/share/apps/mICA/mICA_Toolbox/py/ic_corr.py", line 131, in
value = cor_mat[row][column]
IndexError: list index out of range
To understand this problem, I need more information. Are you running split-half or test-retest reproducibility analysis? If the first, how many repetitions did you choose? if it was only one, this may explain the crash since there is nothing to calculate the correlation coefficients from. Try running more repetitions then. The ICA should not have convergence problems for all samplings. The no_convergence_error.png file should then be in the top folder of the analysis (together with an output "corr_mean.png" that is essentially the figure 2A of our paper). If it is not there, that means that the analysis did not finish properly.

If everything else went smoothly, you can use
python '/Path_to_your_toolbox_installation/py/ic_corr /Path_to_folder_containing_the_samples_folder/ number_of_samplings dimensionality_range 1'
and re-run everything except for the calculation of the single ICAs. Try to leave out the problematic number in in the dimensionality range argument. For example (if you had 10 repetitions):
python '/Path_to_your_toolbox_installation/py/ic_corr /Path_to_folder_containing_the_samples_folder/ 10 2-10,12-20 1'
Let me know if that works.

Best,

Florian
My questions are:
(1) Can I still use this analysis but just ignore the dimensionalities that appear to cause problems?
(2) Why does the groupICA cause problems with dimension 4 but the reproducibility analysis causes a problem with dimension 11? Is this unpredictable as you said?
(2) I unfortunately couldn't find the no_convergence_error.png file - could you give me an indication of where this may be saved?
(3) This may be a silly question but am I correct in assuming that to create a graph such as Figure 2(A) in your paper, I would just need to average the numbers in the corr_.txt files for each dimension?

Thanks very much for your help and sorry for the multiple questions!
Sara
Jun 10, 2016  02:06 AM | Sara De Simoni - Imperial College London
RE: Welcome to Open-Discussion
Dear Florian,

Thank you for getting back to me - apologies if some of my questions were a little vague.

When you say you run melodic, does that mean you run melodic with the same mask, same resolution and smoothing you use in the toolbox? Or do you mean an unmasked ICA? In the latter case it would not be surprising to get completely different results because of different intrinsic dimensionalities of the different brain regions. Unfortunately, we had to remove that part of the results from our toolbox paper prior to publication.

Yes I ran melodic with and without the toolbox in exactly the same way (i.e. with a mask etc). In fact the components I get out are the same but I only get the convergence error when I use the toolbox. In fact the output with the toolbox is all there if I run the groupICA but in the tmp_log I do get the convergence error although it doesn't seem to affect anything. The only reason I noticed was because of the error in the reproducibility analysis.

To understand this problem, I need more information. Are you running split-half or test-retest reproducibility analysis? If the first, how many repetitions did you choose? if it was only one, this may explain the crash since there is nothing to calculate the correlation coefficients from. Try running more repetitions then. The ICA should not have convergence problems for all samplings. The no_convergence_error.png file should then be in the top folder of the analysis (together with an output "corr_mean.png" that is essentially the figure 2A of our paper). If it is not there, that means that the analysis did not finish properly.

I ran the split-half method using the parameters in the attached file. I chose 50 repetitions. My corr_dim.txt files are in the first sample_0001 folder, however I do not get the .png files so I guess something is not working properly. I will definitely try running the last step as you have suggested.


Let me know what you think and thanks again for the help!
Sara


Jun 10, 2016  02:06 AM | Sara De Simoni - Imperial College London
RE: Welcome to Open-Discussion
Hi,

I ran this command as suggested:

python /share/apps/mICA/mICA_Toolbox/py/ic_corr.py /group/HCP/analysis/SaraStriatum/test 50 2-10,12-20 1

However, I still get the following error:

Traceback (most recent call last):
File "/share/apps/mICA/mICA_Toolbox/py/ic_corr.py", line 131, in
value = cor_mat[row][column]
IndexError: list index out of range

I have also uploaded one of the (correct) corr_dim.txt files just to make sure this is what it should contain.
Attachment: corr_dim10.txt