Although I am relatively new to neuroimaging, I have past experience with large datasets and high performance computing and find myself in very strong agreement with the argument in favour of considering HDF5. Indeed, for our own neuroimaging connectivity work, my lab uses a custom (read: hacked together) HDF5-based format.

As one's datasets grow larger, so do the benefits of HDF5; of particular note is the fact that HDF5 works nicely in a parallel distributed-memory computing environment and has been fairly thoroughly tested in that domain. Parallel file I/O is a nontrivial endeavour and it makes life quite a bit easier to have a file format that brings with it a strong library for facilitating this.