ValueError: The size of the proposed random crop ROI is larger than the image size.

[talhaanwarch]

While using

RandCropByPosNegLabeld(
            keys=["image", "label"],
            label_key="label",
            spatial_size=(96, 96, 96),
            pos=1,
            neg=0.5,
            num_samples=4,
            image_key="image",
            image_threshold=0,
        ),

I am getting this error. What parameters should I change to avoid this? like do I need to choose a larger spatial size? Can you also explain what is pos and neg. is it foreground and background?

[Nic-Ma]

Hi @talhaanwarch ,

I think you missed to attach your error.
BTW, pos and neg map to foreground and background.

Thanks.

[talhaanwarch]

I already tried that, but z-axis ranges from 30 to 1000, and padding all images to 1000 to have equal shape is quite expensive. I also noted that notebooks in the tutorial do not use this transform in validation data, but I have to use this, otherwise, I will get dimension error. is it ok to use it?

[Nic-Ma]

OK, maybe @rijobro can help share more advice on the training parameter tuning?

Thanks in advance.

[talhaanwarch]

thanks a lot, I am really having issue in transform selection

[xiefeier]

Dimension Error can be resolved by modifying the parser parameters roi_x, roi_y, roi_z in main.py

[rijobro]

You could use SpatialPadd to ensure your images are all at least (96, 96, 96). This will convert any images with z-size=30 to z-size=96. Any images that were originally bigger than this (e.g., z-size=1000) will remain unchanged. Once all of your images are at least equal to the size of your random crop (96, 96, 96), you can then safely use RandCropByPosNegLabeld.