One thing to keep in mind here is 3dUndump is making a dataset with a single number for each cluster. The number is set by the value placed in the clustcoords.1D serial column. If the clusters overlap, you will only have the last cluster number. This may not be a problem in this particular case because this problem concerns only a minimal size sphere from a group of distinct clusters. But if the clusters were very skinny, then it is possible one cluster could overlap another.

For overlapping clusters, it would probably be better to use 3dUndump to create multiple mask files, a separate mask in each dataset. There is an interesting discussion of part of this problem using 3dcalc from an old message board posting:

[afni.nimh.nih.gov]
[afni.nimh.nih.gov]
[afni.nimh.nih.gov]

Also to avoid calculating the sphere regions outside the brain you will probably want to combine the 3dUndump with the -mask option.