Different outputs for the same transform

[Sebagam]

Hi all,

I'm using two different nifti file datasets, I apply the same transform to them:

Compose([AsChannelFirst(), Resize(spatial_size = [250, 250, 170]), ScaleIntensity(), EnsureType()])

For one group of files I get this tensor:
<class 'torch.Tensor'> torch.Size([2, 1, 250, 250, 170]) tensor([1, 1])

While for the other, I get:
<class 'torch.Tensor'> torch.Size([1, 160, 250, 250, 170]) tensor([1])

What is the issue with the output from the second group and how I can fix it?

Thanks

[Nic-Ma]

Hi @Sebagam ,

Could you please help share your full test program for further analysis?
I feel the data shape you printed out is not expected as channel-first.

Thanks.

[rijobro]

The input to our pre-processing transforms is C, H, W, [D]. It looks like your images do not conform to this, as your first example has a leading [2, 1] and then three spatial dimensions. Your second example has a leading [1, 160], followed by three spatial dimensions.

Resize modifies the images spatial dimensions. You've requested output spatial dimensions of 250, 250, 170. Looking at the output shapes of your examples, they both match in the spatial dimensions.

[Sebagam]

It's [batch size, channels, dim1, dim2, dim3], right?
Do you know why I'm getting 175 channels in the test datasets and how to fix it?
Thanks

[rijobro]

Nifti images don't have batch components, they have the shape x, y, z, t, u, v, w. So I'm not sure what your 2 refers to in your second example. Either way, MONAI doesn't support the batch dimension, and will have to be removed (you could split them into separate images). I would recommend looping over all of your images with something like:

loader = LoadImage()
for filename in list_of_filenames:  # train, validation and test data
    im, meta = loader(filename)
    print(meta["dim"])

It's important to know the variation in shape of your data to choose your transforms accordingly.

c.f.: info on nifti dimensions: https://nifti.nimh.nih.gov/nifti-1/documentation/nifti1fields/nifti1fields_pages/dim.html/document_view