Howdy-

Could we take a step back and check about why an EPI from a subject would be directly aligned to an MNI EPI? In general, EPIs have veeery poor spatial resolution, making high quality alignment tough.

At the risk of repeating something that might already be familiar, I will note that we would recommend:
+ align subject EPI to subject T1w anatomical, say, with just affine alignment because the shapes/structures should be quite similar, different mainly by EPI/other MRI distortions; even though the EPI has poor quality, the idea is that it can be somewhat "easily" be aligned to that subject's own anatomical, which has better detail resolution to use for alignment to a different subject/template brain;
+ align subject anatomical to high definition template, like the MNI T1w template, with nonlinear warping, because both the source and base volumes in this step should have lots of details;
+ concatenate the "subj EPI - subj T1w" and "subj T1w - template T1w" warps (along with any other alignment/registration steps, such as motion correction), and apply one single warp to your original EPI---in this way, the data set isn't interpolated many times, just once, minimizing unnecessary smoothing.

Of course, you might have a specific application in mind, but I thought I would mention the above as a generally preferred modus operandi.

--pt