Dear Krista

Thanks for your information again, you really do me a lot of favor. As your reply, you mentioned that there will have some strange results with using variance normalization during Meta-ICA. Could you tell me more detail about the issue of running META-ICA with variance normalization ? Could you also provide me some figures about "strange results" ?

Many thanks

Best

Paul 

---

Date: Tue, 21 Jan 2014 21:41:30 -0600

From: [log in to unmask]
Subject: Re: [FSL] Different melodic results
To: [log in to unmask]

Hi Chou,

The lower-level ICA set-up to be run 25 times looks fine. I assume your data is already preprocessed and registered to MNI space, and those versions of your data are the ones referenced in subject.txt. We typically use --mmthresh=0.5 (the default) and we do not include --Ostats since we would not be using that extra information. Otherwise this is fine. 

For the higher-level (meta) ICA, you are correct to merge all of the melodic_IC.nii.gz images from the lower-level ICAs together (all remain as unmodified/unthresholded) so your resulting image would be 500 volumes in length if you use -d 20 and run it 25 times. 

We found it useful to include an extra code argument only during the meta-ICA " --vn " which tells MELODIC to switch off variance normalisation. We find this helpful since the data going into the meta-ICA are the components themselves rather than preprocessed data. The outputs can be quite strange (particularly looking at the html of the components) when this is not included. Again, we typically use --mmthresh=0.5 (the default) and we do not include --Ostats since we would not be using that extra information, but that is just how we chose to run it in our lab (we also set the melodics to d60 at both stages). 

Just to note, assuming you want to dual_reg after the meta-ICA, you willof course want to use the unthresholded/unmodified results of the meta-ICA (all components with non removed from the file) as the group-level map/templates for that code. 

Hope that helps.

Best,

Krista

On Mon, Jan 20, 2014 at 10:16 PM, Chou Paul <[log in to unmask]> wrote:

Dear Krista

Based on your helpful suggestions, currently I finished some preliminary simulation for my dataset (this dataset have 400 healthy aged subject, range from 55 - 90 years old). Could you help me to check the following analysis pipeline which I used correct or not ?

1. Lower level ICA

In the first stage, I random select 50 subjects from entire dataset and perform 25 times lower level ICA for different IC components.  I use the following command line for each resample (20 ICs for example):

$FSLDIR/bin/melodic -i subject.txt  -o groupICA -v --nobet --tr=2.5 --report --guireport=../../report.html --bgimage=./2_Overlay_MNI_Template/MNI152_T1_4mm_brain_Resample.nii.gz -d 20 --mmthresh=0.9 --Ostats -a concat

2. Higher level ICA

In the second stage, I combine the "melodic_IC.nii.gz" file of each resampling from lower level ICA using fsl_merge and get the big "All_melodic_IC.nii.gz" file for the Meta_ICA. I use the following command line for the Meta_ICA:

$FSLDIR/bin/melodic -i All_melodic_IC.nii.gz -o MetaICA_IC20 -v --nobet --tr=2.5 --report --guireport=../../report.html --bgimage=../2_Overlay_MNI_Template/MNI152_T1_4mm_brain_Resample.nii.gz -d 20 --mmthresh=0.9 --Ostats

Is the above pipeline reasonable for META-ICA analysis ? Many thanks for your kindly help.

Best

Paul 

---

Date: Mon, 13 Jan 2014 10:14:48 +0800
From: [log in to unmask]

Subject: Re: [FSL] Different melodic results
To: [log in to unmask]

Dear Krista

Thanks for your information, I will try my best to perform analysis according your wonderful suggestions.

Best

Paul 

---

Date: Sun, 12 Jan 2014 11:57:47 -0600
From: [log in to unmask]
Subject: Re: [FSL] Different melodic results
To: [log in to unmask]

Hi Paul,

In terms of the number of "lower-level" MELODICS, each based on a unique subject-order for inputs, we now use 25 now instead of 50 (see the Poppe et al. 2013 paper) since we found that 25 is sufficient. By including 25 instead of say 8 of these random-order "lower-level" MELODICs you are including data from more possible combinations and your results should be more robust. 

Similarly I used random subsets of subjects in a third paper (Wisner KM, Patzelt EH, Lim KO, MacDonald AW 3rd. 2013. An intrinsic connectivity network approach to insula-derived dysfunctions among cocaine users. Am J Drug Alcohol Abuse. Nov;39(6):403-13). The decision to include a random subset of subjects in each "lower-level" melodic was due to computational constraints, which seems to be driven by the resolution of the data at that stage (determined by the resampling resolution) the number of components and the number of subjects. So by using a lower resolution (say 4mm instead of 2mm, the native space of the original data being the best) you may be able to include more subject than say 40 or 50, since lower resolution will require less computations. If you can include more subjects that would probably be best, but it will come down to seeing what your particular system can handle in terms of memory as you go through your analysis. For the NeuroImage, CABN, and AJDAA analyses we ran them on machines requesting 64 GB for the analysis. 

Also, regarding your choice to use 30 components and look at subcomponents of the DMN, at this point we tend to setup the analysis to derive 60 components and with this we get very nice separation of large components into sub components, including the DMN, prefrontal control, limbic, and sensory networks. Additionally, some groups have shown greater sensitivity to group difference at these higher dimensionalities, and based on our current work we find them to be more reproducible than the lower dimensionality (say 20 or 30). So that may be something to consider.

Hope this helps. 

All the best,

Krista

On Sun, Jan 12, 2014 at 8:43 AM, Chou Paul <[log in to unmask]> wrote:

Dear Krista

After quick reading your wonderful papers, I have some questions about the meta-ICA approach and want to hear some feedbacks from you (also the expert from the email list). Currently I am focus on relative large database (400 subjects) and want to perform meta-ICA approach to identify the "stable" sub-component of default mode network (dorsal/ventral and anterior part of DMN) for further dual regression analysis. Would you give me some advise of the following questions ?

I plan to sample 50 subjects from the entire database randomly for 8-15 times and perform MELODIC ICA with 30 ICs for each resample. Do you think 50 subjects is enough for each individual MELODIC ICA?

In your Neuroimage paper, I found you resample/reoder your data 50 times (with 20 ICs for each resample/reorder), but I think this will need a lot of memory for final meta-ICA analysis (the final input of meta-ICA is a single nifti file with 1000 image volumes). I only have a personal computer with 64G memory (I am using centos 5 - 64 bit version), do you think the lower number of resample (only 8-15 times) will cause the instability of identify the DMN subnetwork (or any other network in the brain)?  How many ram do you have for your analysis computer?

Thanks 

Best

Paul

 

 

---

Date: Fri, 10 Jan 2014 10:20:02 -0600
From: [log in to unmask]
Subject: Re: [FSL] Different melodic results
To: [log in to unmask]

Hi Xi,

Others have encountered this problem as well, that is somewhat different MELODIC results depending on the random seed of the algorithm and the order of the subjects. For those worried about the impact of these effects on the reliability of the metrics, some approaches are available. One is a meta-MELODIC and another is ICASSO. Together with colleagues in my lab we tested the meta-MELODIC method and how this play out with regard to reliability in two related papers. They are listed below. 

Wisner KM, Atluri G, Lim KO, Macdonald AW 3rd.

Neurometrics of intrinsic connectivity networks at rest using fMRI: retest reliability and cross-validation using a meta-level method.

Neuroimage. 2013 Aug 1;76:236-51. 

Poppe AB, Wisner K, Atluri G, Lim KO, Kumar V, Macdonald AW 3rd.

Toward a neurometric foundation for probabilistic independent component analysis of fMRI data.

Cogn Affect Behav Neurosci. 2013 Sep;13(3):641-59. 

We continue to pursue this work and are looking into ways to further improve reliability, as well as examining the reliability of different types of metrics derived from meta-MELODIC + Dual Reg results. 

Hope that helps.

Best,

Krista

On Thu, Jan 9, 2014 at 3:26 PM, Tan, Xi <[log in to unmask]> wrote:

Hi, Mark,

Thank you for your reply.  I'm now able to post to the list after being manually added as a subscriber by the helpdesk.

I ran group melodic 6 times, of which 2 have the same subjects order, and the others have different subject orders from the 2 and from each other. The resulting 6 default mode network components, when thresholded by 3, are pretty consistent in precuneus and posterior cingulate region, but vary a lot especially in medial frontal region and right temporal region.

If you would like, I can send you my melodic IC results.

Thank you,
Xi

--

Krista Wisner, M.A. 

Ph.D. Candidate

Clinical Science and Psychopathology Research Program
University of Minnesota
N427 Elliott Hall, 75 E. River Rd.
Minneapolis, MN 55455

--

Krista Wisner, M.A. 

Ph.D. Candidate

Clinical Science and Psychopathology Research Program
University of Minnesota
N427 Elliott Hall, 75 E. River Rd.
Minneapolis, MN 55455

--

Krista Wisner, M.A. 

Ph.D. Candidate

Clinical Science and Psychopathology Research Program
University of Minnesota
N427 Elliott Hall, 75 E. River Rd.
Minneapolis, MN 55455