Transforms not updating meta data

[Spenhouet]

Is your feature request related to a problem? Please describe.

I would have expected that every transform has as input the image and the meta_data dict.
This is not the case for many transforms. They do not update the meta_data.
This leads to inconsistencies and errors since after the transforms the meta_data is no longer valid.

For example:

• AddChannel adds a new channel but does not update the dim, spatial_shape or original_channel_dim in meta_data.
• Spacing returns the new affine but does not update the dim, pixdim, affine or spatial_shape in meta_data.

This is really frustrating. Some transforms are written very well while others are not that great.

Describe the solution you'd like

Please update the meta_data in all transformations. This would also add consistency if every transform expects and returns the meta_data dict.

[rijobro]

Hi @Spenhouet,

I appreciate it is frustrating when functionality isn't quite how you expected/hoped. However, this is a freely-available, community-supported piece of open-source code. Implying that some contributors didn't care about doing things properly is rather unfair and borders on breaching our code of conduct for "showing empathy towards other community members".

I agree that the transforms aren't perfect, but we're working on them! Feel free to add PRs to improve them in line with the changes you'd like to see.

[Spenhouet]

Implying that some contributors didn't care about doing things properly is rather unfair

Sorry, you are right. I updated my comment.

I agree that the transforms aren't perfect, but we're working on them!

That is why I'm opening up this feature request. It seems there is some discussion necessary since not everyone seems to see the meta_data as necessary: #2421
I would vote for making the meta data a first class citizen and including it in all transforms (as parameter and as return value even when there is no change).

[wyli]

I think it is a valid feature request, mainly for the spatial transforms. In fact, we are tracking the changes for the MapTransform using the InvertibleTransform interface

MONAI/monai/transforms/inverse.py

Line 23 in a0afa60

class InvertibleTransform(Transform):

we are not far from this feature after finishing the array class inverse #1749

see also the discussion here #140.