Inversion of background and object in label

[Yangtting]

Hello, I use code below to save predicted label into nii.gz.
When I open the saved “nii.gz" file, I find that the background and objects in the label are completely reversed,
" post_transforms = Compose([
EnsureTyped(keys="pred"),
Invertd(
keys="pred",
transform=test_org_transforms,
orig_keys="image",
meta_keys="pred_meta_dict",
orig_meta_keys="image_meta_dict",
meta_key_postfix="meta_dict",
nearest_interp=False,
to_tensor=True,
),
AsDiscreted(keys="pred", argmax=True, to_onehot=2),
SaveImaged(keys="pred", meta_keys="pred_meta_dict", output_dir="./out", output_postfix="seg", resample=False), ])

    test_data = [post_transforms(i) for i in decollate_batch(test_data)]

"
But I visualize the predicted results using code below, it shows correct.
" test_output = from_engine(["pred"])(test_data)

        original_image = loader(test_data[0]["image_meta_dict"]["filename_or_obj"])[0]

         plt.figure("check", (18, 6))
         plt.subplot(1, 2, 1)
         plt.imshow(original_image[:, :, 168], cmap="gray")
         plt.subplot(1, 2, 2)
         plt.imshow(test_output[0].detach().cpu()[1, :, :, 168])
         plt.show()

"
Also, when I import the saved "nii.gz" file and the original image into ITKSNAP, an error occurs, but I checked that their sizes match.
Could anyone help me?

[Nic-Ma]

Hi @Yangtting ,

Could you please help provide the MONAI version you are using?

Hi @dongyang0122 @ericspod ,

I remember we solved some similar issues before, do you remember more details?
Maybe something related to the image orientation definition.

Thanks in advance.

[dongyang0122]

It might be related to the way how pre-transforms are defined. @Yangtting It would be better if you can share your full transform list.

And the following issue may help here.
#545

[Yangtting]

Here is my code, my MONAI version: 0.8.0, Thank you very much!
test_images = sorted(
glob.glob(os.path.join(data_dir, "images", "*.nii.gz")))

test_data = [{"image": image} for image in test_images]


test_org_transforms = Compose(
    [
        LoadImaged(keys="image"),
        EnsureChannelFirstd(keys="image"),
        Orientationd(keys=["image"], axcodes="RAS"),
        ScaleIntensityRanged(
            keys=["image"], a_min=100, a_max=800,
            b_min=0.0, b_max=1.0, clip=True,
        ),
        CropForegroundd(keys=["image"], source_key="image"),
        EnsureTyped(keys="image"),
    ]
)

test_org_ds = Dataset(
    data=test_data, transform=test_org_transforms)

test_org_loader = DataLoader(test_org_ds, batch_size=1, num_workers=0)

post_transforms = Compose([
    EnsureTyped(keys="pred"),
    Invertd(
        keys="pred",
        transform=test_org_transforms,
        orig_keys="image",
        meta_keys="pred_meta_dict",
        orig_meta_keys="image_meta_dict",
        meta_key_postfix="meta_dict",
        nearest_interp=False,
        to_tensor=True,
    ),
    AsDiscreted(keys="pred", argmax=True, to_onehot=3), 
    SaveImaged(keys="pred", meta_keys="pred_meta_dict", output_dir="./out", output_postfix="seg", resample=False),
])
# # uncomment the following lines to visualize the predicted results
from monai.transforms import LoadImage
loader = LoadImage()
device = torch.device("cuda:0")
model = UNet(
    dimensions=3,
    in_channels=1,
    out_channels=3,
    channels=(16, 32, 64, 128, 256),
    strides=(2, 2, 2, 2),
    num_res_units=2,
    norm=Norm.BATCH,
).to(device)
os.chdir(savepath)

model.load_state_dict(torch.load(modelpath))
model.eval()

with torch.no_grad():
    for test_data in test_org_loader:
        test_inputs = test_data["image"].to(device)
        roi_size = (160, 160, 160)
        sw_batch_size = 4
        test_data["pred"] = sliding_window_inference(
            test_inputs, roi_size, sw_batch_size, model)

        test_data = [post_transforms(i) for i in decollate_batch(test_data)]

[dongyang0122]

@Yangtting the pre-transforms looks good. But I notice that you have set "to_onehot" to 3 in the post-transform. You can try to set it as None to see if it works. When "to_onehot" is not None, it will create additional dimension for classes.

https://docs.monai.io/en/stable/transforms.html#monai.transforms.AsDiscreted

[Yangtting]

Thank you for your reply, when I set ”to_onehot“ to None, my bug was solved!

When I trained the model before, I set it to 3(two targets and one background), it works well. So why I coudn't set it to 3 in testing process? Could you please explain it for me? Thanks a lot.

[dongyang0122]

For model training, loss functions require one-hot ground truth in general. But when saving the prediction, one-hot predictions are not necessary for visualization as the one-hot operation would create a 4D matrix instead of a 3D one.

[Yangtting]

Thank you! You are really nice!

[rafiibnsultan]

Thank you! Was searching for this for 2 days