Hi @contsili,
Agree with what Francois said. It's okay to use a subset of ROIs (probably wise). Whether or not you can see anything with 50 trials depends on the SNR for your specific signal. It doesn't seem like very many trials to me, but I work mostly with really teeny signals.
Although connectivity measures are fancy and many people are using them with MEG, I am finding them a bit too variable between measures and parameter-dependent for my liking, and that is even before we go to unconstrained / subcortical / deep sources. We are exploring those issues in some known signals before we'll be relying on them too much. I doubt that your expected connectivity pattern is going to jump out of your data.
You may find that a combination of theory/literature driven highly restricted connectivity questions and/or looking at basic things like which ROI signal peaks first might be a stronger and more convincing way to go for now and might still allow you to speculate on connectivity in that system.
Since you're working on cerebellum, this paper might be useful for your general reference: https://www.sciencedirect.com/science/article/pii/S1053811920303049
Good luck,
Emily