saveimaged saving 2D masks in nifti format flipped

[OmarAshkar]

Describe the bug
I am trying to save masks from hovernet pipeline similar to the tutorial here: https://github.com/Project-MONAI/tutorials/blob/main/pathology/hovernet/inference.py. When reloading the same masks by nibabel.load(), I get the images flipped (y,x) instead of (x,y), then flipped up.

I have to have a 32bit format to capture instance segmentation mask, so png didn't work for me and tif produces large files, so Nifti is what I need now.

To Reproduce
Steps to reproduce the behavior:

1. run the hovernet pipeline from the tutorial for link above
2. load one of the outputs by nibabel.load(), then .get_fdata()

Expected behavior
It should have been in the same dimensions of original image.

Additional context
This is crucial in this context, as png format supports only uint8, and tif files are huge and cannot be compressed from saveimaged.

Thanks!

[wyli]

could you please provide an example of the saving/loading?

[OmarAshkar]

Hi @wyli Thank you for your response. But I still didn't get it. I have images that I feed to this script and the outputs are flipped on saving nifti. This is not the case if I use PIL save as png. If you can elaborate please would be appreciated. Thanks!

[KumoLiu]

Hi @OmarAshkar, I just tried the Hovernet inference script and didn't find x,y has been flipped. Could you please share more information on how to find x,y has been flipped?

import matplotlib.pyplot as plt
import nibabel as nib

pre_transforms = mt.Compose(
    [
        mt.LoadImaged(keys="image", reader=PILReader, converter=lambda x: x.convert("RGB")),
        mt.EnsureChannelFirstd(keys="image"),
        mt.CenterSpatialCropd(keys="image", roi_size=(900, 800)),
        mt.CastToTyped(keys="image", dtype=torch.float32),
        mt.ScaleIntensityRanged(keys="image", a_min=0.0, a_max=255.0, b_min=0.0, b_max=1.0, clip=True),
    ]
)

out = pre_transforms({"image":  image_path})
data = nib.load(inference_instance_map_path)

print(out["image"].shape)  # --> torch.Size([3, 900, 800])
print(data.get_fdata().shape) # --> (900, 800)

Thanks!

[OmarAshkar]

Hi @KumoLiu Thank you for putting the time to look into that. To reproduce the error using the HoverNet tutorial, simply open one of the saved nii masks and compare it visually to the corresponding images.

Also, I think image shape is the same. It is just rotated in a certain way.

Thanks!

[KumoLiu]

Hi @OmarAshkar, yes I just open one of the saved predictions and compare it visually to the corresponding images. They are not rotated.

print(out["image"].shape)  # --> torch.Size([3, 900, 800])
print(data.get_fdata().shape) # --> (900, 800)
plt.subplot(1,2,1)
plt.imshow(out["image"].permute(1,2,0))
plt.subplot(1,2,2)
plt.imshow(data.get_fdata())

Thanks!

[OmarAshkar]

@KumoLiu I don't know what am I missing here. I will try to check my transformers to see which one causes that. But I can repreduce it from my end by switching ".nii" to ".png", things work as expected.