Evaluation with CropForegroundd

[mikiwang820]

Hello,

I used "CropForegroundd" in the validation transform when I was training the model, so I still applied it when I did evaluation and saved my predictive labels.
However, it caused the outputs had the different sizes from the ground truth.
Therefore, I take it out, and the outputs' sizes are right, but the dice score drop from 0.6 to 0.4.

How can I deal with this problem.
Thank you for any help you can provide

[rijobro]

You should be able to apply the same transform to both your ground truth and your segmentation. That way they'll both be the same size. Something like this:

CropForegroundd(keys=["image","label"], source_key="label")

Here, both image (your ground truth) and the label (your segmentation) will be cropped using a lambda (by default voxels >0) applied on the source_key image (label).

[rijobro]

Check this tutorial for an example of CropForegroundd.

[FrancescoLR]

Hello, I am joining the thread with a related question.
What about if after inference I want to pad the output to the original input size (as before CropForegoundd)? Is there a straightforward way of doing this?
Thanks a lot.

[rijobro]

The coordinates used for CropForegroundd are stored in data["image_meta_dict"]["foreground_start_coord"] and data["image_meta_dict"]["foreground_end_coord"] by default. You could get these values and then pad by that amount using BorderPad.

Alternatively, I'm working on being able to perform the inverse of transforms, which would return the image to the original size. Feel free to try out the PR and use this tutorial PR for hints along the way.