With regard to special treatments, obviously you should avoid different preprocessing strategies within a study. However, it might be better to accept a potential bias due to excluding the first few volumes in one subject, thus shortening the session, instead of relying on crappy volumes (or replacing those by other volumes).
Anyway. I think you should have a look at the realignment parameters. How do they look like when realigning onto the first volume, how do they look like when realigning onto the mean, what happens when excluding the first few volumes. Maybe the changes are very small anyway.