How about creating spheres of whatever size at each cluster center of mass as your mask? Would that work for you? Do this for each cluster. If you want to make it more sophisticated, the spheres might be masked against the clusters too and possibly count the voxels also. You could progressively increase the size of the sphere until you get to the number of voxels you would like roughly.
3dcalc includes this example in the help.
7. Create a region-of-interest mask comprised of a 3-dimensional sphere.
Values within the ROI sphere will be labeled as '1' while values
outside the mask will be labeled as '0'. Statistical analyses can
then be done on the voxels within the ROI sphere.
The example below puts a solid ball (sphere) of radius 3=sqrt(9)
about the point with coordinates (x,y,z)=(20,30,70):
3dcalc -a anat+tlrc \
-expr 'step((9-(x-20)*(x-20)-(y-30)*(y-30)-(z-70)*(z-70))'\
-prefix ball