The conventional whole-brain voxel-wise group analysis starts with the presumption that the neighboring voxels are unrelated in the group (as well as the individual subject) model. To compensate for the unrealistic presumption, a second step is typically required: correction for multiple testing. Such a splitting strategy can be over-penalizing and inefficient as evidenced by the difficulty of having some clusters pass the narrow funnel of FWE controllability.
If the group analysis is performed on a list of ROIs (defined through atlas, parcellation, or prior information), an alternative approach is to build one holistic model for those ROIs instead of hundreds of thousands of separate models for the voxels in the brain. Such an integrative model is more efficient (and less penalizing) since all the ROIs are treated equally in terms of effect strength without discrimination over their anatomical size.
A program called BayesianGroupAna.py is available now for running such a ROI-based group analysis. The underlying mechanism is laid out in the manuscript https://www.biorxiv.org/content/early/2018/02/20/238998. Read the help for more details:
Let me know if you have questions, feedback, suggestions, etc.