ResizeWithPadOrCropd does not pad background channel in labels.

[umgpy]

Describe the bug
ResizeWithPadOrCropd does not pad background channel for label data.

Expected behavior
For a channel wise binary data, with background as 1, ResizeWithPadOrCropd should add 1's instead of 0's where padding takes place

Screenshots

Top row is the ground truth label, with padding and bottom (predicted labels) is the expected padding behavior.

Environment

Ensuring you use the relevant python executable, please paste the output of:

================================
Printing MONAI config...
================================
MONAI version: 0.7.0
Numpy version: 1.21.1
Pytorch version: 1.7.1+cu92
MONAI flags: HAS_EXT = False, USE_COMPILED = False
MONAI rev id: bfa054b9c3064628a21f4c35bbe3132964e91f43

Optional dependencies:
Pytorch Ignite version: 0.4.6
Nibabel version: 3.2.1
scikit-image version: 0.18.2
Pillow version: 8.3.1
Tensorboard version: 2.6.0
gdown version: NOT INSTALLED or UNKNOWN VERSION.
TorchVision version: 0.8.2+cu92
tqdm version: 4.62.3
lmdb version: NOT INSTALLED or UNKNOWN VERSION.
psutil version: NOT INSTALLED or UNKNOWN VERSION.
pandas version: 1.3.1
einops version: 0.3.2
transformers version: NOT INSTALLED or UNKNOWN VERSION.

For details about installing the optional dependencies, please visit:
    https://docs.monai.io/en/latest/installation.html#installing-the-recommended-dependencies


================================
Printing system config...
================================
`psutil` required for `print_system_info`

================================
Printing GPU config...
================================
Num GPUs: 4
Has CUDA: True
CUDA version: 9.2
cuDNN enabled: True
cuDNN version: 7603
Current device: 0
Library compiled for CUDA architectures: ['sm_37', 'sm_50', 'sm_60', 'sm_70']
GPU 0 Name: Tesla V100-DGXS-32GB
GPU 0 Is integrated: False
GPU 0 Is multi GPU board: False
GPU 0 Multi processor count: 80
GPU 0 Total memory (GB): 31.7
GPU 0 CUDA capability (maj.min): 7.0

Additional context
My code for transforms-

fn_keys = ["Img","Seg"]


val_transforms = Compose(
    [
        LoadImaged(keys=fn_keys),
        EnsureChannelFirstd(keys="Img"),
        ConvertToMultiChanneld(keys="Seg"),
        ResizeWithPadOrCropd(spatial_size=(256,512,512),keys=fn_keys),),
        EnsureTyped(keys=fn_keys),
    ]
)

[Nic-Ma]

Hi @umgpy ,

I think you can set the constant arg of np.pad as 1 for padding in the transform:
https://github.com/Project-MONAI/MONAI/blob/dev/monai/transforms/croppad/dictionary.py#L1424

Thanks.

[wyli]

perhaps change the padding mode to "edge" https://github.com/Project-MONAI/MONAI/blob/dev/monai/transforms/croppad/dictionary.py#L1416-L1417 (I'm converting this to a discussion)

[umgpy]

Hi, @Nic-Ma , @wyli

Thank you for the quick replies!

Trying constant_values args doesn't work as it fills the value for all channels, as seen here-

ResizeWithPadOrCropd(spatial_size=(256,512,512),keys=fn_keys, constant_values = 1)

At the same time mode '"edge"' for the label does work channel wise but it has unintended consequence of elongating the label mask, which is not right.

 ResizeWithPadOrCropd(spatial_size=(256,512,512),keys=fn_keys, mode = ['constant','edge'])

Thanks again for all the help!

[umgpy]

I also tried "median" which works mostly

But at times there are artifacts, such as this-

[wyli]

I see, perhaps you can do spatial padding before ConvertToMultiChanneld

[umgpy]

Thanks, this solved it!