ICA-AROMA droped the 32/44 components as noise for subject with low head motion. Is that right?

Hi, I use ica-aroma of fmriprep to denoise data. But I got a worse result.
The result becomes fragmented !!!

As of now, ICA is not helpful. I try to figure out possible problems:

  1. I found denoise algorithm drop the too many components. All components is 44, but dropped noise components is 32. But Head motion is fine. Mean FD is 0.057.
    The situation happen for many subjects.

  2. Denoise algorithm misclassify the signal and noise?

  1. I thought ICA-AROMA already have do the 6 mm smoothing, so I didn’t do smoothing in the 1stLevel analysis (generally, I use 8mm smoothing). Does it matter? or Actually,I need do it ?

Hello,

What are these results of? What did you expect the results to look like and what other denoising methods have you compared to?

More components being noise compared to signal is normal.

I don’t think this is misclassified.

The AROMA outputs are smoothed with a 6mm kernel. There’s no one right answer in terms of which smoothing kernel works best. In part it is motivated by how much spatial specificity you need vs SNR reduction you can tolerate. That is, if you are interested in a small region, like an amygdala sub-component for example, you would not want a big smoothing kernel, even though SNR might be reduced.

If you’d prefer to use a different kernel or a different space, you can make a non-aggressive AROMA-denoised output yourself using FSL:

fsl_regfilt -i sub-<subject_label>_task-<task_id>_space-<space>_desc-preproc_bold.nii.gz \
    -f $(cat sub-<subject_label>_task-<task_id>_AROMAnoiseICs.csv) \
    -d sub-<subject_label>_task-<task_id>_desc-MELODIC_mixing.tsv \
    -o sub-<subject_label>_task-<task_id>_space-<space>_desc-AROMAnonaggr_bold.nii.gz

Best,
Steven

This is my previous result without ICA(same fmriprep process, 8mm smoothing). It is a group result which covary subjects’ accuracy. I put head motion regressors into GLM to denoise data.