Choosing a statistical model for a 2X4 mixed fMRI design (and implementing it in FSL's FEAT)

Hi everyone :wave:!
I am trying to come up with the most suitable statistical model to test my hypotheses on an fMRI neuroimaging study, and its correct implementation in FSL’s FEAT. The design is as follows:

Subjects (N=68) participated in four fMRI sessions. During each session they performed a neurofeedback (NF) task repeatedly (1-4 NF runs per session), learning to upregulate a brain region from which they recieved sensory feedback (total of 9-11 NF runs along 4 session per subject). Subjects where randomized into one of two groups (34 subjects in each). The groups differed only in the targeted brain region from which they recieved the feedback (Test vs Control target regions).

I have performed a standard 1st level analysis (on 672 NF runs, mind you!) on FSL’s FEAT, and I wonder how to continue. I hypothesize about differences between groups (Test vs Control) and sessions in specific brain regions, which I will approach with ROI analysis and a mixed linear model. Additionally, I would like to inspect whole brain effects of time (a linear regressor of sessions) and groups. Specifically, I am interested in contrast maps of last vs first session for both groups seperately, and contrast maps of the difference between last and first session, between both groups (i.e. Test(last vs first)>Control(last vs first)).

I wonder whether to conduct a 2nd level analysis, averaging each session (each containing 1-4 runs) in a fixed effects design (assuming I do not care about the within session variance and would like to have robust estimates of activity per session), and then devise a 2X4 repeated measures ANOVA model: with one between-subjects factor (Test/Control) and one within-subjects repeated measure variable (sessions). If this makes sense, I am not sure how to build such a model in FSL. I went through FEAT’s documentation and no such model were explained (where you have 1 beween-subjects factor and one repeated measures factor with 4 levels).

Does this description make sense? How would I build it in FEAT’s model setup? How would I treat subjects in this model?

Alternatively, I may build two separate models for each group (Test and Control) in order to produce the within group contrast maps, and create contrasts of interest between groups based on models of the specific levels of interest (e.g. 4th vs 1st sessions. I actually do not have a general group hypothesis across all sessions, only specific simple effects such as difference in the 4th session between groups). The issue with this solution is that I do not go through the general test of interaction effects before looking into simple effects.

Thanks in advance,
Nitzan.