Hi @rijobro, this PR is mainly to set deepcopy=False here:

Hm, although disabling deep copying will make things faster, we obviously need to be careful of accidentally modifying an image's meta data when modifying another image that is a shallow copy. I suppose id(self) != id_in is limited in that with a = b, both a and b have different ID's, but they are shallow copies of one another and so here a shallow copy would suffice.

I'm not sure what the easy answer here is, we need a cleverer mechanism for figuring out when to copy and when not.

I don't think this is necessarily a good idea, but we could have a context manager:

with MetaTensor.deep_copy_meta(False):
    out = model(image

/integration-test

I don't think this is necessarily a good idea, but we could have a context manager:

with MetaTensor.deep_copy_meta(False):
    out = model(image

I'm trying to set shallow copy when is_batch is True, essentially when it's a batch, it's within a network forward and the meta info are not updated within it.

I suppose when it's in a batch we're unlikely to be modifying meta data. Seems like a bit of an indirect correlation, though.

the existing unit/integration tests work fine and the changes are non-breaking, I'll merge this for now but please feel free to create better solutions.