channels_last doesn't work

🐛Bug

Hi, I’m learning TorchIO and I found channels_last doesn’t do anything with an input tensor (C, W, H, D).

To reproduce

import numpy as np
import torchio as tio

print(f'{tio.__version__=}')

# Input is numpy array
array = np.random.rand(1, 256, 256, 159).astype('float32')
print('\nInput numpy array')
print(array.shape)

print('\ntio.ScalarImage with channels_last=False')
array_ScalarImage = tio.ScalarImage(tensor=array, channels_last=False)
print(array_ScalarImage.shape)

print('\ntio.ScalarImage with channels_last=True')
array_ScalarImage = tio.ScalarImage(tensor=array, channels_last=True)
print(array_ScalarImage.shape)

Expected behavior

I expected it to return the tensor (D, C, W, H) according to the document.

Actual behavior

tio.__version__='0.18.39'

Input numpy array
(1, 256, 256, 159)

tio.ScalarImage with channels_last=False
(1, 256, 256, 159)

tio.ScalarImage with channels_last=True
(1, 256, 256, 159)

System info

Output of python <(curl -s https://raw.githubusercontent.com/fepegar/torchio/master/print_system.py):

Platform: Linux-5.4.0-73-generic-x86_64-with-glibc2.10
TorchIO:  0.18.39
PyTorch:  1.7.1
NumPy:    1.20.3
Python:   3.8.5 (default, Sep  4 2020, 07:30:14)
[GCC 7.3.0]

[romainVala]

yes we should add it more explicitly in the doc…

Yes the extra batch dimension is added by pytorch dataloader, and not in the torchio transform

The specific ordering of the spatial dimension (D, H, W) does not really matter, for all the transform … whatever the spatial ordering is you get the same at the end of the transform … it is more a matter for the network inputs, to ensure the image are always presented similarly for MRI data, we use the transform ToCanonical in order to have all data slice in the same spatial axis …

[romainVala]

Hello

this is not a bug.

channels_last argument, is only used when your read image from disk ( nifti file) . In nifti 4D file, the 4D dimension is often the last dimension. If we want this last dimension to be interpreted as the channel dimension, the argument channels_last = True, will just permute the 4 dimension, to be first (channel dimension in pytorch)

If you use tensor as input, then no need of this extra argument, just reorder your tensor first …