ExtractDataKeyFromMetaKeyd transform is intended to always return a dictionary and not modify the MetaTensor itself.

Yes this is true but I should restate with an example. This transform, like others that have the "d" suffix, are dictionary transforms which accept dictionaries and inputs and produce them as outputs, both of which represent a set of named values such as an image with its label. If we load an image called "foo" from "foo.nii" the input to LoadImaged will be {"foo": "foo.nii"}. Doing so with image_only=False will produce an output dictionary where "foo" maps to the loaded image tensor and "foo_meta_dict" maps to the metadata dictionary for that image.

ExtractDataKeyFromMetaKeyd allows us to pull a value from "foo_meta_dict" and add it to our data dictionary as a new key. This value can then be used in other transforms. The change that I see to do is to allow this transform to reference a MetaTensor instead of this dictionary and pull a value from its metadata instead.

This is the example using metadata dictionaries:

inputs={"foo": "foo.nii"}  # initial input mapping item name to path

# load the data using LoadImaged
li = LoadImaged(image_only=False)
dat = li(inputs)   # dictionary with keys "foo" (image tensor) and "foo_meta_dict" (metadata dict)

# use transform to pull value "filename_or_obj" in "foo_meta_dict"
e = ExtractDataKeyFromMetaKeyd("filename_or_obj", meta_key="foo_meta_dict")
dat = e(dat)  # now has keys "foo", "foo_meta_dict", and "filename_or_obj"

# verify that the data was pulled correctly from the nested dictionary
assert dat["foo_meta_dict"]["filename_or_obj"] == dat["filename_or_obj"]

Instead I want to be able to refer to the MetaTensor itself and pull the same piece of data from its meta dictionary:

inputs={"foo": "foo.nii"}  # initial input mapping item name to path

# load the data using LoadImaged
li = LoadImaged() 
dat = li(inputs)   # dictionary with keys "foo" (image tensor)

# use transform to pull value "filename_or_obj" in metadata of "foo"
e = ExtractDataKeyFromMetaKeyd("filename_or_obj", meta_key="foo")
dat = e(dat)  # now has keys "foo" and "filename_or_obj"

# verify that the data in the MetaTensor metadata matches what we pulled out
assert dat["foo"].meta["filename_or_obj"] == dat["filename_or_obj"]

I think this is a feasible change to make such that the transform can detect if the meta_key argument references a dictionary or a MetaTensor, then behaves accordingly. If you wanted to start work on this and coordinate later I will leave it up to you.

Enhance ExtractDataKeyFromMetaKeyd to work with MetaTensor #7562

Description

Metadata

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Issue actions

Enhance `ExtractDataKeyFromMetaKeyd` to work with `MetaTensor` #7562