Hi @Uri_Shinitsky ,
It really depends on your ROI size, and balance between tolerance for spatial bluriness vs SNR goals. So, no single solution. If you have really small ROIs (e.g., a hippocampal subfield), or are interested in neighboring areas and would not want to blur the signal between them, then you probably want small smoothing kernel. If you have a low SNR acquisition and have large and/or distant ROIs, then you can probably afford a larger smoothing kernel.
A good place to start is looking at similar publications, or trying out different smoothings yourself on an a separate independent (but ideally similar) dataset.
Best,
Steven