Issue with Volume-Segmentation Mismatch after Using MetadataCopier

[Nastii22]

Hello @amine0110 ,

First of all, I want to thank you for your excellent documents on medical imaging.

While I was using nnUNet, I encountered errors and warnings during preprocessing before using "MetadataCopier." After using "MetadataCopier," everything worked fine with nnUNet training. The issue I faced is that when I checked the volume and corresponding segmentation in both ITK-SNAP and 3D Slicer before using "MetadataCopier," they matched. However, after using "MetadataCopier" (and also after training and checking inference), they no longer matched. I attached the images of 3D Slicer before and after using "MetadataCopier" to better illustrate the issue.
I would appreciate your assistance.

[amine0110]

Hi @Nastii22

Where are you using the metadatacopier?

I have never used it!

[Nastii22]

Hi @amine0110 , Thank you for getting back to me.

I watched your YouTube video on Vertebrae Segmentation using Python and nnUNetv2 (https://www.youtube.com/watch?v=rZfngRIhEcQ). Around the 45-46 minute mark, you mentioned the importance of ensuring that the metadata in both scans and segmentations (annotations) match. You explained that this can be achieved by copying the metadata from the scans and pasting it into the segmentations. You also referred to code example on your GitHub repository: https://github.com/amine0110/spine-segmentation-nnunet/blob/main/prepare.py.

I encountered some errors and warnings during preprocessing:
RuntimeError: Some images have errors. Please check text output above to see which one(s) and what's going on.
Warning: Affine is not the same for image and segmentation!
Error: Spacing mismatch between segmentation and corresponding images.
Warning: Direction mismatch between segmentation and corresponding images.
Warning: Origin mismatch between segmentation and corresponding images.

Based on your comment in the YouTube video, I used the "Metadatacopier" before uploading my data and providing it to nnUNet. After using "MetadataCopier," everything worked fine with nnUNet training, except for the issue I mentioned previously.

Thank you in advance.

[amine0110]

Hum, this is strange, I didn't encounter this from my end.

But I think something happened with the volumes.

What you can try, is to run the inference on the original volume before running it to the metadatacopier, and then try to overlay the output on the original volume.

Please try this and let me know what you have got and we can see from there.

[Nastii22]

Hi @amine0110 ,

The problem was related to my dataset orientation. I corrected it and now everything looks fine.

Thank you so much for your time and assistance.