fix: apply transpose for voxel_spacing

[thangngoc89]

Edit:
This is a more complex problem then I initially thought so I converted this into a draft.
Summary of what I have learned so far:

• SimpleITK loads NifTI and store data in ZYX order. This is the same order that you get from SimpleITK.GetArrayFromImage
• SimpleITK returns spacing in physical world, meaning XYZ. Reference from nnunet
• nnUNet preprocessor expect the input data and spacing both in ZYX order.
• By applying the transpose_backward order, you're turning nnunet's plan spacing to ZYX order.
• The code still misses the part where you need to give SimpleITK resampler the XYZ spacing order.
• A good place to add this is in moose function

desired_spacing = desired_spacing[::-1]

I feels like the problem has become too complex and I'm testing interaction between multilevel of transforming.

---

Transposed axes in nnUNet experiments place the largest spacing first. Voxel spacing in plans.json is stored in transposed order, so a backward transpose is needed to restore the original ZYX order when using SimpleITK.

Reference for nnUNet experiment planner:
https://github.com/MIC-DKFZ/nnUNet/blob/3df680c6d28ea88735fca1c1921df380471c1af6/nnunetv2/experiment_planning/experiment_planners/default_experiment_planner.py#L215-L226

[vercel]

The latest updates on your projects. Learn more about Vercel for Git ↗︎

Name	Status	Preview	Comments	Updated (UTC)
moose	✅ Ready (Inspect)	Visit Preview	💬 Add feedback	Jan 15, 2025 6:27pm