Less number of slices in the training data.

[ramchandracheke]

Hi,
I am training a model with input 128x128x64, However, few samples have less than 64 slices in the training. Can you please suggest how to handle this problem.
Thanks,
Ram

[ramchandracheke]

train_transforms = Compose(
    [
        LoadImaged(keys=["image", "label"]),
        AddChanneld(keys=["image", "label"]),
        ScaleIntensityRanged(
            keys=["image"], a_min=-200, a_max=300,
            b_min=0.0, b_max=1.0, clip=True,
        ),
        SpatialPadd(keys=["image", "label"],spatial_size=(128,128,64),method= 'symmetric'),
        RandCropByPosNegLabeld(
            keys=["image", "label"],
            label_key="label",
            spatial_size=(128, 128, 64),
            pos=1,
            neg=1,
            num_samples=4,
            image_key="image",
            image_threshold=0,
        ),
        ToTensord(keys=["image", "label"]),
    ]
)

[rijobro]

Can you remove the RandCropByPosNegLabeld and confirm that all images are at least (128,128,64)? The documentation says:

if a dimension of the input data size is bigger than the pad size, will not pad that dimension

[ramchandracheke]

Hey @rijobro,
Thanks for reply. Some of the CT scan data contains less than 64 slices. Before, I did not apply below transformation.

SpatialPadd(keys=["image", "label"],spatial_size=(128,128,64),method= 'symmetric'),

But after adding this transformation it works. It is okay to add this order of transformation?
Thanks,
Ram

[rijobro]

Yeah, if you do SpatialPadd followed by RandCropByPosNegLabeld, then you're essentially saying "I want all of my images to be at least (128,128,64), and then if any are bigger, I want you to take a random crop, such that the resulting image is (128,128,64).