The problem seems to improve when using high-res masks (1mm isotropic). In this case only a smaller fraction of seed-voxels is affected. Seems like there occurs a rounding error at some point. Nevertheless strange that it never occurs when using only single-voxel masks.