Dear FSL experts,

I've recently finished pre-processing (using DPARSFA) resting-state fMRI data from ~250 subjects with varying diagnoses (principally depression, MCI and Alzheimer's disease). The majority of these data are from the PsyMRI consortium, and thus the data that I've been processing are derived from around 10 different scanners/sites.

In brief, I plan to use spectral DCM to investigate whether there is a common brain network across diagnostic groups.

In the past I have used stochastic DCM to examine effective connectivity in depression where each node in the model was defined by specific components following ICA using melodic (i.e., using dual regression approach to extract subject-specific timecourses from peak voxels across ICA maps).

I would ideally like to pursue a similar method using the current multi-site data, but am unsure of the best approach to model (and ideally remove) between scanner/site variation. I've canvassed the opinions of several colleagues, who have suggested a range of options (e.g., controlling for site post-ICA; using the ComBat method on end-point connectivity metrics)...it seems, however, that there is no gold standard approach.

I am aware that there are a number of approaches that have been tested (e.g., FIX, ICA-AROMA, linked-ICA [but this appears to be for multimodal approaches?]), but wasn't sure whether any of these are able to control for site-specific variables (e.g., scanner make/model; scan length; differences in TR), or whether (in the case of ICA-AROMA) they only model head motion?

Given the nature of the data, it's unfortunately not possible to balance the different diagnostic groups across different sites (e.g., some sites provided only MCI cases, others only MDD), which I will try to address when analysing group effects on the networks.

I guess the key questions are:

1. Would group ICA or PCA (i.e., using melodic; perhaps MIGP) be an adequate approach to both capture common resting-state networks across sites, whilst handling spurious site-specific variables??

2. Alternatively, would ICA-AROMA or FIX be a more suitable option? I can see that the latter is being used in the UK Biobank analysis pipeline, but am aware that the scanners across sites are the same make/model - do these methods extend to different scanner manufacturers/models?

2a. Assuming FIX was a good option, what would be the best way to train the classifier given I've got ~20-30 subjects/site? (does it need to be trained for each site, or would a set number/site, say 5/site, suffice, which can then be applied to the rest of the data)?

3. Is it possible to use either, or all, of the approaches above with data with differing TRs? (I understand GIFT allows this, but wasn't sure about the FSL tools)

4. I read on another post that using dual regression on ICA maps (derived from melodic) from data with different scan lengths may be fraught, especially when it comes to comparisons (https://www.jiscmail.ac.uk/cgi-bin/webadmin?A2=FSL;5e397cb1.1609); is there a valid or known workaround for modelling subject-specific ICA timecourses of different lengths?

Sorry in advance for the long post - I'd be grateful for any input you might have :)

Best wishes,
Matt

########################################################################

To unsubscribe from the FSL list, click the following link:
https://www.jiscmail.ac.uk/cgi-bin/webadmin?SUBED1=FSL&A=1