Dear Raimon and Maarten,

Thank you very much for the input. I agree it is a bit of stretch to evaluate the performance by the number of retained components. Nevertheless, for some patients, it was retaining around 5~6 out of 40~50, which made me a bit worried at that time.

It turns out that as an input file to AROMA, I was feeding in MNI-registered image, along with affmat and warp files; the original low-res functional file was warped twice as a result, which totally produced garbage results (many of the ICs were also mapped outside of the brain). What I am doing right now is skipping the register2MNI function in the AROMA script. With this, more than twice as many components are retained and all the ICs exist within the brain. As for the preprocessing step, I am following the instructions of delaying the temporal filtering until after AROMA.

Just out of curiosity, is the temporal filtering delayed until the end so that the HF feature of the classifier won't be biased during classification?

Once again, thank you very much.

On Tue, Jul 14, 2015 at 7:53 AM, Maarten Mennes <[log in to unmask]> wrote:
Dear Andrew,

just based on the number of components that is retained it is hard too make the conclusion that it is throwing away too much data. You would have to assess the quality of the remaining data (as we did in our manuscripts) to decide whether AROMA selected good data to be removed. Afterall it is possible that you have several noise sources in your data that all are 'powerful' enough to end up in separate ICA components.

One way to assess would be to run a groupICA on the cleaned data across participants. Does it still come up with noise components after applying the denoising?

Maarten



On Tue, Jul 14, 2015 at 12:30 PM, Pruim, R.H.R. (Raimon) <[log in to unmask]> wrote:
Dear Andrew,

20-30% seems indeed a bit much (we retained ~11 out of ~35 components in our training sample; http://dx.doi.org/10.1016/j.neuroimage.2015.02.064 ), but it is difficult to say if the classification is actually going wrong here. Can you share the full command you're running? Also, what kind of data do you have (e.g. TR, #time-points, voxel-size) and which preprocessing did you run before ICA-AROMA? It might also be worth checking the raw/input data and some of the melodic output (melodic.ica/report/00index.html) to check if there are any notable issues with the data.

Best,
Raimon






-----Original Message-----
From: FSL - FMRIB's Software Library [mailto:[log in to unmask]] On Behalf Of Andrew Song
Sent: vrijdag 10 juli 2015 16:20
To: [log in to unmask]
Subject: [FSL] ICA-AROMA throwing out too many components

Dear FSL experts,

I have tried running ICA-AROMA on some of the resting state datasets that I have.

It seems that, for each subject, the algorithm seems to be retaining only around 20~30% (~ 10 out of total ~40 component0 of the ICs and regards rest as motion/noise components. Although it could be that dataset is very noisy, to me it seems that the algorithm is throwing away too much data.

Do you have any insights as to why this might be happening?

Thanks,



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




--
Andrew Hyungsuk Song
MIT Class of 2015
Electrical Engineering & Computer Science
617.999.8077