Transform without labels in segmentation

[talhaanwarch]

When we want to predict on test data without labels (mask), how to transform in that case.

I think it should be something like this

train_transforms = T.Compose(
            [
        T.LoadImaged(keys=["image"]),
        T.AddChanneld(keys=["image"]),
        T.Spacingd(keys=["image"],pixdim=(1.5, 1.5, 2.0),mode=("bilinear", "nearest")),
        T.ScaleIntensityd(keys=["image"]),
        T.CropForegroundd(keys=["image"], source_key="image"),
        T.RandCropByPosNegLabeld(
            keys=["image"],
            label_key="label",
            spatial_size=(64, 64, 64),
            pos=1,
            neg=1,
            num_samples=8,
            image_key="image",
            image_threshold=0,
        ),
        T.ToTensord(keys=["image"]),
            ]
        )

But how CropForegroundd and RandCropByPosNegLabeld can know that its foreground and its pos neg label without having a label?
One way is to remove these two transforms.

I removed RandCropByPosNegLabeld as in spleen tutorial before removing this transform, my result looks like (a bit acceptable).
(The following show transform with both image and label)
The first image is ground truth, the second inverted ground truth, third is predicted in each row.

After removing RandCropByPosNegLabeld transform techniques results become horrible like this

[rijobro]

Cropping the foreground is something you can only do if you have the ground truth. So at training time, you can use them, but at validation/testing time, you'll probably want to use a sliding window. In this way, you'll still process images that are the same size as they were at training time, but now you can do the whole image. The spleen segmentation tutorial has examples of this.

[talhaanwarch]

I followed the spleen tutorial, and used sliding_window_inference but you can see that results become very bad when I am not using RandCropByPosNegLabeld during inference

[talhaanwarch]

do CropForegroundd also need ground truth?

[Nic-Ma]

Hi @talhaanwarch ,

CropForegroundd transform uses the select_fn arg to determine foreground:
https://github.com/Project-MONAI/MONAI/blob/dev/monai/transforms/croppad/dictionary.py#L803

Thanks.

[talhaanwarch]

issue of bad prediction after RandCropByPosNegLabeld is because of using small roi of (32,32,32). Making it (160,160,160) fixed the issue