Print

Print


Hi all, I'm currently applying all the new knowledge I got from this year's FSL course. I have questions regarding using ARMOA on single-subject task MRI to remove noise components. The FSL course/practical had a very useful tutorial on looking at/identifying noise and motion components manually and with FIX and AROMA. However it didn't contain a full example on how to incorporate these into a task-based analysis, so could I get some feedback/guidance on whether the strategy I've inferred is correct/reasonable?

To confirm, the general strategy is:

(Step 1) Load in original 4D data, run a First-level analysis: Preprocessing (in FEAT GUI) and do everything I would be planning to do in the full analysis (pre-stats, and registration)

(Step 2) Use whatever method (AROMA, FIX, manual) to ID components that will be removed

(Step 3) Use fsl_regfilt to remove these components from the original 4D data

(Step 4) Load in this new denoised data and run a First-level analysis: Statistics (in FEAT GUI) .. continue as normal with the stats and post-stats

I've been reading about ICA-AROMA in the provided papers (both very helpful) but I am still a big confused about the correct steps for implementing it as no one in my lab has taken this strategy yet. The manual suggests leaving the MELODIC work to the AROMA command; it also suggests not doing temporal filtering at that stage. So if I understand this correctly, if I'm planning to run ICA-AROMA after FEAT the specific strategy I would do is:

(Step 1) Load in original 4D data, run a First-level analysis: Preprocessing (in FEAT GUI) and do everything I would be planning to do in the full analysis (pre-stats and registration) . 

Question: in pre-stats, (a) don't turn on MELODIC, (a) turn off BET, and (a) turn off temporal filtering. Is that correct? Do I also (4) change "High pass filter cutoff" in the "Data" tab to something else so it can be applied later - e.g. after doing the full model so I can estimate the high pass filter from the design? (Or is this completely ignored until the post-processing stage anyway?)

(Step 2) Run AROMA by specifying the feat directory and the output. 

Question: it's not clear to me how I can incorporate a mask here (since using FEAT's extraction is not advised for AROMA). If I'm running AROMA on a feat directory is this step taken care of for me (as the manual states "MELODIC will be run within ICA-AROMA using the appropriate mask" - that means it will do the bet for me on the example_func image or should I be doing a bet on this image myself before running AROMA)?

(Step 3) Use fsl_regfilt to remove these components from the original 4D data after inspecting the identified components in fsleyes.

Question: Before I move on to step 4, is it necessary to do the highpass temporal filtering by doing another preprocessing run with only highpass temporal filtering (everything else unchecked)? Based on the Pruim et al (2015) paper Fig 1 it looks like the temporal filtering was done as a step between what I describe as steps 3 & 4.

(Step 4) Load in this new denoised data and run a First-level analysis: Statistics (in FEAT GUI) .. continue as normal with the stats and post-stats. 

Thank you for your feedback. If I've completely missed resources somewhere that has all this information, please let me know!

Cheers,
Kim