Hello,

Thank you very much for your reply, Maarten. Just to confirm, the optimal option would be:
1) apply Melodic to the 5 runs (not concatenated, a list)
2) concatenate temporally all the runs (fslmerge -t) to obtain a subject_image

3) apply fsl_regfilt -i subject_image -o subject_image_denoised -d melodic_mix -f "1,4"
4) split the subject_image_denoised into 5 runs_denoised, to continue analyzing each run separately 

Is this option correct?

I ask because I am not sure if makes sense not concatenating in Melodic but concatenating in fsl_regfilt.
I made a proof concatenating the 5 runs for the input of Melodic, setting automatic ICs, and the resulted ICs were only 3 and all bad ICs. Instead, if the input is the 5 runs without concatenation, 15 automatic ICs led to good and bad ICs.

Best,

Úrsula



2015-11-06 9:04 GMT+01:00 Maarten Mennes <[log in to unmask]>:
Hello Ursula,

if I follow your reasoning I doubt option B makes sense given that you seem to be removing specific components from each run. However, realize that the components you are removing were estimated across all runs. Therefore removing some components from some runs, but not others would be inappropriate. Optimal in option B would be to run fsl_regfilt across your concatenated runs. 

hth
Maarten

On Thu, Nov 5, 2015 at 9:19 PM, Úrsula Pérez <[log in to unmask]> wrote:
Hi all,

I am participating in a resting-state fMRI study. Before group ICA, I would like to remove some non-interest independent components for each subject (5 runs with 90 volumes each one), applying Melodic and fsl_regfilt. Each run has been preprocessed separately.

I am considering two options, for each subject:

A) Melodic for each run, obtain ICs and fsl_regfilt for each run 
    (5 times because there are 5 runs)

B) Melodic with a list of the 5 runs as input and obtain ICs.
    Then, apply fsl_regfilt to each run. For that, it is necessary to split melodic_mix in matrices of 90 rows (because each run has 90 volumes), i.e.
fsl_regfilt -i run1 -o run1_denoised -d melodic_mix_rows1to90 -f "1,2"
fsl_regfilt -i run2 -o run2_denoised -d melodic_mix_rows91to180 -f "1,5"
fsl_regfilt -i run3 -o run3_denoised -d melodic_mix_rows181to270 -f "8,2"
fsl_regfilt -i run4 -o run4_denoised -d melodic_mix_rows271to360 -f "7,2"
fsl_regfilt -i run5 -o run5_denoised -d melodic_mix_rows361to450 -f "9,2"

I prefer option B because it requires less workload to identify good or bad components, and also because maybe one run is not enough for ICA.

Please, could you tell me if option B is correct?

Thank you very much,
All the best,

Úrsula



--
Maarten Mennes, Ph.D.
Senior Researcher
Donders Institute for Brain, Cognition and Behaviour
Radboud University Nijmegen
Nijmegen
The Netherlands