I agree with @Victor_G . That is how you should do it.
It depends on the nature of the data, of course, but removing 1 to 3 observations might not make a big difference to the D-efficiency. You may also need to think about augmenting with new runs to get a more useful data set.