MONAI dataloader producing samples of inconsistent dimension

[philnovv]

Hi all, here is my set of transforms I'm using:

[
           LoadImaged(
               keys=["image", "label"],
               allow_missing_keys=True),
           EnsureChannelFirstd(
               keys=["image", "label"],
               allow_missing_keys=True),
           Orientationd(
               keys=["image", "label"],
               axcodes="RAS"),
           Spacingd(
               keys=["image", "label"],
               pixdim=config["voxel_size"],
               mode=("bilinear", "nearest")),
           NormalizeIntensityd(
               keys=["image"],
               channel_wise=True),
           CastToTyped(
               keys=["label"],
               dtype=(torch.int16)),
           SpatialPadd(
               keys=["image", "label"],
               spatial_size=config["train_patch_size_voxels"]),
           RandCropByPosNegLabeld(
               keys=["image", "label"],
               label_key="label",
               spatial_size=config["train_patch_size_voxels"],
               pos=config['pos_sample_ratio'],
               neg=1,
               allow_smaller=False)
       ])

Using this set of transforms, I sometimes obtain image samples with different dimensions, e.g.
[1,92,92,92] and [1,92,92,87].

My question is: how is this possible? My understanding is that if the image size is less than e.g. 92 in any dimension, then SpatialPadd will pad it to 92 where necessary. Since this happens before random crop, shouldn't the image always have dimensions greater than or equal to the crop dimensions?

A second issue: despite calling "allow_smaller=False" in the random cropping init, no error is raised my MONAI when the inconsistent samples are produced.

Any thoughts would be greatly appreciated!

[KumoLiu]

Hi @philnovv, you should move the SpatialPadd after RandCropByPosNegLabeld or add another SpatialPadd.

how is this possible? My understanding is that if the image size is less than e.g. 92 in any dimension, then SpatialPadd will pad it to 92 where necessary. Since this happens before random crop, shouldn't the image always have dimensions greater than or equal to the crop dimensions?

Yes, SpatialPadd will ensure all your output larger or equal to 92. But RandCropByPosNegLabeld will random generate the crop center which can not ensure your output still have 92.

A second issue: despite calling "allow_smaller=False" in the random cropping init, no error is raised my MONAI when the inconsistent samples are produced.

allow_smaller is used to check whether your data is smaller than the roi size before crop since you have SpatialPadd. So after it, all the image size will >= 92 then will not raise exception.

MONAI/monai/transforms/croppad/dictionary.py

Line 919 in e1a69b0

allow_smaller: if `False`, an exception will be raised if the image is smaller than

Hope it helps, thanks.