What you describe is already what the connectivity processes do in Brainstorm, when you select "Apply scout function: Before" with unconstrained source maps in input. There is no need to develop anything more.
In @Raymundo.Cassani's experiments, we could observe that computing the average per ROI before computing the connectivity measure ("Scout function: before") degrades the quality of the estimation significantly. The best results seem to be obtained by computing the connectivity graph between all the individual vertices involved in all the ROIs, and then take the maximum across all the connectivity values between two ROIs. However, this becomes impractical for full-brain connectomes, as it requires computing matrices that can't fit in the memory of any regular computer (15000x15000xNfrequencies).
We are still actively discussing this topic, this is why our documentation takes so long to be published... We hope to get some clear recommendations in the next few weeks.