The goal of this aggressive masking procedure isn’t to remove null differences, but rather to account for differences in brain coverage across the images in the analysis. The contrast maps being used in the contrast permutation meta-analysis should be unthresholded, so non-significant values should still be present in the maps. In unthresholded maps, values of zero are more likely to indicate voxels outside the brain mask than estimates of exactly zero.
Unfortunately, none of NiMARE’s IBMA methods can currently handle missing data (e.g., a voxel at the edge of the brain where some of the maps have valid values and others do not), so we chose to just ignore those voxels from the start. For a bit of a discussion about this, you can see https://github.com/neurostuff/NiMARE/issues/274.
I hope that helps.