RandRotated

[FabianHeu]

I would like to rotate my image label pairs with the RandRotated function to improve robustness:

RandRotated(keys=["image", "label"], prob=1.0, range_x=1, range_y=1, range_z=1)

But it ends up that it is rotating the image differently to the label, which doesn't make sense and screws up all the labeling. Is this a planned behaviour or am I missing something?

Help appreciated

Fabian

[FabianHeu]

Ok found the error in the plotting. It's the same problem than in #1788
But I dont't really understand why this is not working

[rijobro]

If your issue is the same as in #1788, then the effect you're seeing is because each time you call RandRotated, it will apply a new random rotation to the data. So if you call it twice, once for image, and then again for label, then they will have different rotations. If you just call it once with:

transform = RandRotated(keys=["image", "label"], ...)
out_data = transform(in_data)
rotated_im, rotated_label = out_data["image"], out_data["label"]

then the same random rotation will have been applied to both the image and the label.

[rijobro]

If it still doesn't work, please share a minimum working example.

[FabianHeu]

it works now but I still don't get it. This is how it works

training_transform = Compose([
    LoadImaged(keys=['image', 'label']),
    AddChanneld(keys=["image", "label"]),
    RandRotated(keys=["image", "label"], prob=1.0, range_x=1, range_y=1, range_z=1),
    ToTensord(keys=['image', 'label']),`

training_dataset = PersistentDataset(data=training_data_dict, transform=training_transform)

plot_train = training_dataset[0]                                                                                                    # here is the difference
figure1 = plt.figure(num="After_Pipeline", figsize=(8, 4))
plt.suptitle("Image label pair of {}. training-dataset with name: {}".format(i + 1, str3))
plt.subplot(1, 2, 1)
plt.title("image")
plt.xlabel(str1)
saggital_image = plot_train["image"][0, 60, :, :]                                                                        # here is the difference
plt.imshow(saggital_image, cmap="gray")
plt.subplot(1, 2, 2)
plt.title("label")
plt.xlabel(str2)
saggital_label = plot_train["label"][0, 60, :, :]                                                                            # here is the difference
plt.imshow(saggital_label, cmap="gray")
plt.show()

An this how it didnt work:

figure1 = plt.figure(num="After_Pipeline", figsize=(8, 4))
plt.suptitle("Image label pair of {}. training-dataset with name: {}".format(i + 1, str3))
plt.subplot(1, 2, 1)
plt.title("image")
plt.xlabel(str1)
saggital_image = training_dataset[0]["image"][0, 60, :, :]                                                   # here is the difference
plt.imshow(saggital_image, cmap="gray")
plt.subplot(1, 2, 2)
plt.title("label")
plt.xlabel(str2)
saggital_label = training_dataset[0]["label"][0, 60, :, :]                                                       # here is the difference
plt.imshow(saggital_label, cmap="gray")
plt.show()

[rijobro]

Each time you call training_dataset[0], they will be randomly rotated by a different amount. In the top example, you call training_dataset[0] once, but in the bottom example you call it twice. Therefore, in the bottom example, saggital_image and saggital_label have two different rotations applied to them.

[FabianHeu]

Thanks a lot! I really appreciate your help. And by the way this framework and your work is really awesome!!