I’m doing an ICA study on the ADNI dataset, which contains two types of rs-fMRI data: single-echo (TR 3 s) and multi-echo (TR 0.607 s). Is it okay to use these two types of images in the same analysis? I think that if I apply low-pass filtering of at least < 0.33 Hz the differences in TR shouldn’t matter, or am I thinking too simply? Do I need to apply certain transformations in the data (e.g., downsampling multi-echo data by choosing every other 5 volumes), or should I exclude the multi-echo images altogether (similar to this paper)?
In that paper you linked, it looks like they excluded multiband fMRI data, rather than multi-echo data. Based on the TRs you mentioned, multiband makes more sense than multi-echo. Am I interpreting that correctly?
If you are talking about multiband data, then I would say that it depends on how you plan to employ ICA. What approach do you plan to use?
Thanks for your reply. Yes, I should’ve said multiband and not multi-echo, as it is also how the ADNI data description has specified it.
Briefly, I’m going to first preprocess the data using fmriprep, and then use FSL MELODIC and dual regression to compare the salience network between two groups of patients. The templates for dual regression will be based on group-level ICA performed on the same subjects (from both groups).
Thank you for the clarification. It might be good to change the title and tags for the topic based on that.
I can’t claim to be an expert on this (I mostly responded because I do a reasonable amount of multi-echo fMRI), but my understanding of dual regression is that it involves a time-concatenated ICA to estimate the group-level components. I do not think that temporally filtering your multiband data will accomplish what you need, as the concatenated time series used for the group-level ICA will contain multiple sampling rates.
The idea of downsampling the multiband data sounds promising, although I think it makes sense to interpolate rather than decimate the data. I haven’t come across any similar approaches in the literature, though, so hopefully someone else will be able to weigh in on it.
Depending on the sample sizes, I would look more deeply into temporal downsampling as a possibility. Alternatively, you could run this as two analyses- essentially an analysis and a replication. As long as your hypotheses are well-defined, this is a good opportunity to see how robust your findings are.
Thanks for your explanation. Yes, it doesn’t make sense to concatenate images with different sampling rates in the group-level ICA. Using the multiband data in a replication analysis is a great idea, but I’m not sure if the sample size in single-/multi-band data is large enough to do it. So I should look more into the best approach of doing downsampling in this case.
As for the title and tags I tried to edit them in the first place, but I don’t seem to have the permission to do it. I would appreciate it if moderators can change the title and tags.