Hi, Dong-
Yes, this is a classic case of "identifying brain/nonbrain looks easy to the eye, but the computer doesn't quite get it". Whether this will have deep affects on further analysis, depends on what you are trying to do. And note that 3dAutomask is often meant to be a quick-and-convenient approximation, and not a quantitative brain masking approach. So, what are you aiming to do?
Re. ways of improving/tweaking 3dAutomask: You could use the "-clfrac .." option, such as "-clfrac 0.6". Perhaps "-peels 2" might help. But note that these kinds of refinements can be very scanner-by-scanner and subject-by-subject.
--pt