Possible bug .nrrd vs nifti incorrect transformations

[evaherbst]

Hi all,

We discovered a bug a long time ago (fall 2024) when creating a new segmentation model.
Unfortunately that student left and I cannot share the data to replicate the issue since it was medical imaging data but I wanted to note the behaviour here in case anyone else finds it useful.
The bug was as follows: .nrrd files caused issues with the back transformation of the inferred segmentation.
First we thought the model had not been trained properly, but we realized that the segmentation was specific to the subject/scan but distorted and in the wrong position (not aligned with image). We hypothesize that a back transform into original imaging space (after presumable alignment used for the inference) was not applied correctly.

Using .nifti files solved the issue so we migrated our workflow to that.