Resampled Image Gives Unexpected Shape

[slefcourt27]

Hello,

I am working with MRI data and have an example image which has shape (384, 300, 32) with pixdim [0.5208333 0.5208333 4.8]. Resampling the image with a transform = tio.Resample((1, 1, 4.8)) gives a new image with shape (192, 150, 7). I expect the depth to remain the same at 32. Why is it now 7?

Similarly, if I resample that same image to 1mm isotropic, the output shape is (192, 150, 16). Why is the depth now 16? I would expect it to be ~154.

I appreciate any help.

[fepegar]

Can you please share your image and some code to reproduce?

…

On Tue, 13 Feb 2024, 05:02 Samuel L., ***@***.***> wrote: Hello, I am working with MRI data and have an example image which has shape (384, 300, 32) with pixdim [0.5208333 0.5208333 4.8]. Resampling the image with a transform = tio.Resample((1, 1, 4.8)) gives a new image with shape (192, 150, 7). I expect the depth to remain the same at 32. Why is it now 7? I appreciate any help. — Reply to this email directly, view it on GitHub <#1144>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/ADAZVVGM4IHYE2YKJNR7HLTYTLXWXAVCNFSM6AAAAABDF3PKKKVHI2DSMVQWIX3LMV43ERDJONRXK43TNFXW4OZWGIYTEMJRGQ> . You are receiving this because you are subscribed to this thread.Message ID: ***@***.***>

[slefcourt27]

I'm not able to share the image I am working with. But, I was able to reproduce one of the issues using the braTS dataset. Please see the attached ZIP file. Thank you for your help!

!pip install torchio

import torchio as tio
import nibabel as nib
import numpy as np
import nibabel as nib
from pathlib import Path

t2 = Path('./BraTS20_Training_001_t2.nii')
mask = Path('./BraTS20_Training_001_seg.nii')
filenames = [t2, mask]

nifti_img = nib.load(filenames[0])
nifti_mask = nib.load(filenames[1])
header = nifti_img.header

img = nifti_img.get_fdata()
mask = nifti_mask.get_fdata()

img = np.rot90(img, 3)
mask = np.rot90(mask, 3)

print('Image shape', img.shape)
print('Pixdim', nifti_img.header['pixdim'])

img = np.expand_dims(img, axis=0).copy()
mask = np.expand_dims(mask, axis=0).copy()

subject = tio.Subject(
    image = tio.ScalarImage(tensor=img),
    mask = tio.LabelMap(tensor=mask)
)

# Now replicating my use case.
transform = tio.Resample((0.5208333, 0.5208333, 4.8), image_interpolation='linear', label_interpolation='nearest')
transformed_image = transform(subject.image)
print('Original Shape', transformed_image.shape)

# First experiment - should maintain the depth while width and height are divided by 2.
# This ends up working (but not in my use case?)
transform2 = tio.Resample((1, 1, 4.8), image_interpolation='linear', label_interpolation='nearest')
result = transform2(transformed_image)
print(result.shape)

# Second experiment - depth should increase by a factor of ~5.
# This does not work as expected.
transform3 = tio.Resample((1, 1, 1), image_interpolation='linear', label_interpolation='nearest')
result = transform2(transformed_image)
print(result.shape)

Tochio_Reproduce.zip

[fepegar]

Can you try reading the images with TorchIO instead of NiBabel? Have you gone through the tutorials?

[fepegar]

>>> import numpy as np
>>> import torch
>>> import torchio as tio

>>> tensor = torch.rand(1, 100, 100, 100)
>>> spacing = 0.5208333, 0.5208333, 4.8
>>> affine = np.diag((*spacing, 1))
>>> image = tio.ScalarImage(tensor=tensor, affine=affine)
>>> image
ScalarImage(shape: (1, 100, 100, 100); spacing: (0.52, 0.52, 4.80); orientation: RAS+; dtype: torch.FloatTensor; memory: 3.8 MiB)

>>> resample = tio.Resample((1, 1, 4.8))
>>> resample(image).shape
(1, 53, 53, 100)

>>> resample = tio.Resample(1)
>>> resample(image).shape
(1, 53, 53, 480)

[slefcourt27]

Thank you for your help! I had not previously gone through the colab tutorials. Not reading the images using TorchIO was part of the issue. I was also switching between numpy and TorchIO Image. Staying consistent with TorchIO fixed it completely.