Apply transforms to label that has more than one channel

[Alberto1404]

Hello everyone. I am still a beginer at MONAI, and I am facing a problem I don't know how to solve.
To define the train and validation transforms for creating the data loaders, I am doing as follows:
from monai.transforms import ( EnsureChannelFirstd, Compose, LoadImaged, Orientationd, RandFlipd, RandRotated, RandZoomd, NormalizeIntensityd, ToTensord )
The issue comes when dealing with my label. Unlike other common situations, where the shape is (H,W,D) and each voxel takes one scalar, my label has a shape of (3,H,W,D), where each voxel is a array of three values. When applying the transformations (either training or validation), I am getting this warning:
Loading dataset: 0%| | 0/23 [00:00<?, ?it/s]/home/guijosa/Documents/PythonDocs/PythonDocs_env/lib/python3.8/site-packages/monai/data/utils.py:536: UserWarning: Modifying image pixdim from [1. 1. 1. 0.] to [ 1. 1. 1. 363.37824646] warnings.warn(f"Modifying image pixdim from {pixdim} to {norm}") /home/guijosa/Documents/PythonDocs/PythonDocs_env/lib/python3.8/site-packages/monai/data/utils.py:536: UserWarning: Modifying image pixdim from [1. 1. 1. 0.] to [ 1. 1. 1. 364.24922786] warnings.warn(f"Modifying image pixdim from {pixdim} to {norm}") /home/guijosa/Documents/PythonDocs/PythonDocs_env/lib/python3.8/site-packages/monai/data/utils.py:536: UserWarning: Modifying image pixdim from [1. 1. 1. 0.] to [ 1. 1. 1. 363.54023161] warnings.warn(f"Modifying image pixdim from {pixdim} to {norm}") /home/guijosa/Documents/PythonDocs/PythonDocs_env/lib/python3.8/site-packages/monai/data/utils.py:536: UserWarning: Modifying image pixdim from [1. 1. 1. 0.] to [ 1. 1. 1. 364.37686535] warnings.warn(f"Modifying image pixdim from {pixdim} to {norm}") Loading dataset: 4%|████▊ | 1/23 [00:02<00:53, 2.44s/it]/home/guijosa/Documents/PythonDocs/PythonDocs_env/lib/python3.8/site-packages/monai/data/utils.py:536: UserWarning: Modifying image pixdim from [1. 1. 1. 0.] to [ 1. 1. 1. 363.94195966] warnings.warn(f"Modifying image pixdim from {pixdim} to {norm}") /home/guijosa/Documents/PythonDocs/PythonDocs_env/lib/python3.8/site-packages/monai/data/utils.py:536: UserWarning: Modifying image pixdim from [1. 1. 1. 0.] to [ 1. 1. 1. 366.44576679] warnings.warn(f"Modifying image pixdim from {pixdim} to {norm}") /home/guijosa/Documents/PythonDocs/PythonDocs_env/lib/python3.8/site-packages/monai/data/utils.py:536: UserWarning: Modifying image pixdim from [1. 1. 1. 0.] to [ 1. 1. 1. 365.27215881] warnings.warn(f"Modifying image pixdim from {pixdim} to {norm}") /home/guijosa/Documents/PythonDocs/PythonDocs_env/lib/python3.8/site-packages/monai/data/utils.py:536: UserWarning: Modifying image pixdim from [1. 1. 1. 0.] to [ 1. 1. 1. 363.651619] warnings.warn(f"Modifying image pixdim from {pixdim} to {norm}") Loading dataset: 35%|██████████████████████████████████████▉ | 8/23 [00:04<00:06, 2.37it/s]/home/guijosa/Documents/PythonDocs/PythonDocs_env/lib/python3.8/site-packages/monai/data/utils.py:536: UserWarning: Modifying image pixdim from [1. 1. 1. 0.] to [ 1. 1. 1. 364.00206043] warnings.warn(f"Modifying image pixdim from {pixdim} to {norm}") /home/guijosa/Documents/PythonDocs/PythonDocs_env/lib/python3.8/site-packages/monai/data/utils.py:536: UserWarning: Modifying image pixdim from [1. 1. 1. 0.] to [ 1. 1. 1. 364.50720157] warnings.warn(f"Modifying image pixdim from {pixdim} to {norm}") Loading dataset: 39%|███████████████████████████████████████████▊ | 9/23 [00:06<00:09, 1.55it/s]/home/guijosa/Documents/PythonDocs/PythonDocs_env/lib/python3.8/site-packages/monai/data/utils.py:536: UserWarning: Modifying image pixdim from [1. 1. 1. 0.] to [ 1. 1. 1. 364.06283798] warnings.warn(f"Modifying image pixdim from {pixdim} to {norm}") Loading dataset: 48%|█████████████████████████████████████████████████████ | 11/23 [00:06<00:05, 2.18it/s]/home/guijosa/Documents/PythonDocs/PythonDocs_env/lib/python3.8/site-packages/monai/data/utils.py:536: UserWarning: Modifying image pixdim from [1. 1. 1. 0.] to [ 1. 1. 1. 364.77595864] warnings.warn(f"Modifying image pixdim from {pixdim} to {norm}") /home/guijosa/Documents/PythonDocs/PythonDocs_env/lib/python3.8/site-packages/monai/data/utils.py:536: UserWarning: Modifying image pixdim from [1. 1. 1. 0.] to [ 1. 1. 1. 365.05547524] warnings.warn(f"Modifying image pixdim from {pixdim} to {norm}") Loading dataset: 57%|██████████████████████████████████████████████████████████████▋ | 13/23 [00:07<00:05, 2.00it/s]/home/guijosa/Documents/PythonDocs/PythonDocs_env/lib/python3.8/site-packages/monai/data/utils.py:536: UserWarning: Modifying image pixdim from [1. 1. 1. 0.] to [ 1. 1. 1. 380.58080614] warnings.warn(f"Modifying image pixdim from {pixdim} to {norm}") /home/guijosa/Documents/PythonDocs/PythonDocs_env/lib/python3.8/site-packages/monai/data/utils.py:536: UserWarning: Modifying image pixdim from [1. 1. 1. 0.] to [ 1. 1. 1. 380.89598318] warnings.warn(f"Modifying image pixdim from {pixdim} to {norm}") Loading dataset: 70%|█████████████████████████████████████████████████████████████████████████████▏ | 16/23 [00:08<00:03, 2.15it/s]/home/guijosa/Documents/PythonDocs/PythonDocs_env/lib/python3.8/site-packages/monai/data/utils.py:536: UserWarning: Modifying image pixdim from [1. 1. 1. 0.] to [ 1. 1. 1. 364.84483003] warnings.warn(f"Modifying image pixdim from {pixdim} to {norm}") /home/guijosa/Documents/PythonDocs/PythonDocs_env/lib/python3.8/site-packages/monai/data/utils.py:536: UserWarning: Modifying image pixdim from [1. 1. 1. 0.] to [ 1. 1. 1. 384.86068908] warnings.warn(f"Modifying image pixdim from {pixdim} to {norm}") /home/guijosa/Documents/PythonDocs/PythonDocs_env/lib/python3.8/site-packages/monai/data/utils.py:536: UserWarning: Modifying image pixdim from [1. 1. 1. 0.] to [ 1. 1. 1. 379.04056511] warnings.warn(f"Modifying image pixdim from {pixdim} to {norm}") Loading dataset: 74%|██████████████████████████████████████████████████████████████████████████████████ | 17/23 [00:10<00:04, 1.39it/s]/home/guijosa/Documents/PythonDocs/PythonDocs_env/lib/python3.8/site-packages/monai/data/utils.py:536: UserWarning: Modifying image pixdim from [1. 1. 1. 0.] to [ 1. 1. 1. 382.50718947] warnings.warn(f"Modifying image pixdim from {pixdim} to {norm}") Loading dataset: 87%|████████████████████████████████████████████████████████████████████████████████████████████████▌ | 20/23 [00:11<00:01, 2.31it/s]/home/guijosa/Documents/PythonDocs/PythonDocs_env/lib/python3.8/site-packages/monai/data/utils.py:536: UserWarning: Modifying image pixdim from [1. 1. 1. 0.] to [ 1. 1. 1. 381.05445805] warnings.warn(f"Modifying image pixdim from {pixdim} to {norm}") /home/guijosa/Documents/PythonDocs/PythonDocs_env/lib/python3.8/site-packages/monai/data/utils.py:536: UserWarning: Modifying image pixdim from [1. 1. 1. 0.] to [ 1. 1. 1. 382.83645333] warnings.warn(f"Modifying image pixdim from {pixdim} to {norm}") /home/guijosa/Documents/PythonDocs/PythonDocs_env/lib/python3.8/site-packages/monai/data/utils.py:536: UserWarning: Modifying image pixdim from [1. 1. 1. 0.] to [ 1. 1. 1. 380.11248335] warnings.warn(f"Modifying image pixdim from {pixdim} to {norm}") Loading dataset: 100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████| 23/23 [00:12<00:00, 1.92it/s]

The 4D volume I am saving uses the nifti affine and the nifti header from its correspondant input volume.
Thank you for your time :)

[rijobro]

If you do image, meta = LoadImage()(label) where label is the filename of one of your label images, what do you get for image.shape, and what is the contents of meta?

[rijobro]

Hi, looks fine to me. What seems to be the problem?

[Alberto1404]

In my case I am working with a network that has more than one decoder. Reason why working better with dictionary transforms. I am working in a binary scenario (out_channels = 1). The first decoder is the typical semantic segmentator (each value in Ground-Truth takes value 0-background or 1-foreground). When applying transforms here, I get a normal pytorch tensor with (batch_size, channels, H,W,D). Nevertheless, for this second decoder, where my ground-truth (the commented label above) has that 4D-shape, as it is not possible to work with 6D pytorch tensors (batch, channels, ?, H,W,D) I though to directly consider those 3 as literal channels (although we are in a binary scenario). When doing this, I got that warning message applying the transforms to this label.
I don't know whether it will affect to the predictions.

[rijobro]

If you are doing binary segmentation, why does your ground truth label consist of 3 channels? What does each channel represent?

You mentioned that you have 6D tensors, but I think you have 4 channels: C,H,W,D with shape 3,512,512,96.

Some ideas:

• Maybe your label is RGB, and each channel dimension is identical to the others?
• Maybe your label contains 3 labels? Background, label 1 and label 2?

Feel free to send your code across and I'll have a look.

[Alberto1404]

In that label, each channel represents coordinates x,y and z of vectors located at each voxel position.

[rijobro]

Ok, so your label is a deformation or displacement field? And your network is trying to infer that deformation? Is this not just non-rigid registration?