2d not axial affine

[lilipj]

Hi,
I've 2d slices nifti files, generated from 3d with nibabel. These files are not axial, I've sagittal or coronal data.
The 3d affine is correct. I would like to extract the 2d affine with the to_affine_nd() function.
I wonder if this function works when data are not in axial orientation, because it takes the top left 2x2 elements of the 3d affine,
but I think it's only for axial orientation. Is that correct ?
Thanks in advance

[wyli]

yes, that utility function always extracts the upper left 2x2, you can normalise the spatial orientation with Orientationd before the extraction.
`

MONAI/monai/transforms/spatial/dictionary.py

Lines 393 to 395 in 422cc6d

	class Orientationd(MapTransform, InvertibleTransform):
	"""
	Dictionary-based wrapper of :py:class:`monai.transforms.Orientation`.

[wyli]

in my understanding, the file you uploaded "patient_1.nii.gz" is a 3d volume with shape (512,512,1), or do I misunderstand:

>>> print(nibabel.load("patient_1.nii.gz").header)
<class 'nibabel.nifti1.Nifti1Header'> object, endian='<'
sizeof_hdr      : 348
data_type       : b''
db_name         : b''
extents         : 0
session_error   : 0
regular         : b'r'
dim_info        : 0
dim             : [  3 512 512   1   1   1   1   1]
intent_p1       : 0.0

[anais2390]

Oh you are right, I am sorry I uploaded the wrong file. Actually, the code I posted correspond to what append when I load a 2d file (512, 512), I will provide you with the good file !
Actually, uploading a 3d file (512, 512, 1) is the solution I found to avoid the problems with to_nd_affine()

patient_1_slice.nii.gz

[wyli]

so this file is not in a good shape, and can't be read by 3d slicer/itk-snap:

if it's generated automatically by MONAI components, I'm curious about how it's generated and it's probably a bug from MONAI...

[wyli]

for future reference, given the input slice (512, 512, 1) in 3D (patient_1.nii.gz), the 2d analysis could be done by:

import torch

from monai.data import *
from monai.transforms import *
from monai.utils import first


def main():
    train_dict = [{"image": "patient_1.nii.gz"}]

    train_transforms = Compose(
        [
            LoadImaged(keys="image", image_only=True, ensure_channel_first=True, as_closest_canonical=True),
            SqueezeDimd(keys="image", dim=2, update_meta=True),
            Spacingd(keys="image", pixdim=(5, 5)),
        ]
    )

    train_ds = CacheDataset(data=train_dict, transform=train_transforms)
    train_loader = DataLoader(train_ds, batch_size=1, shuffle=True, pin_memory=torch.cuda.is_available())

    check = first(train_loader)
    print(check["image"].shape, check["image"].affine)


if __name__ == "__main__":
    main()

the as_closest_canonical=True will make it more robust when SqueezeDim drops any of the spatial dimensions.

[anais2390]

Thanks you @wyli !
The 2d file was not generated via MONAI but using nibabel librairy.