Thanks. I tried with -multi_cost so I could compare different cost functions. Some are better, but none are great, and -nmi and -lpa actually make the anatomical quite distorted (to use a technical term, it looks squished).

Not sure if this is playing a role whatsoever, but I also looked at the @SSwarper output again, and noticed in the posterior portion of the brain it is not very precise-looking (not sure how to describe this, see attached image). I'm wondering if it's having to do a lot of work to try to make it fit the template since the field of view wasn't great in the original acquisition, and if this might be affecting alignment to the EPI as well.