SpatialResample vs ResampleToMatch vs Spacing

[hermancollin]

Hi. I'm looking into the resampling transforms to get a common pixel size (e.g. 0.1 um/px) for all my 2D images across training/validation/test splits. I'm confused about the distinction between these 3 transforms:

• SpatialResample
• ResampleToMatch
• Spacing
  Looking at the source code, it seems the last 2 are related to the SpatialResample transform.

I understand an integral part of these transforms is the affine matrix inside the image metadata. How does this work for 2D PNG images, for example? From the load_medical_images tutorial, I can see the LoadImaged transform will read a PNG image using PILReader and give a 4x4 identity matrix as affine matrix. How can I embed the pixel size of my images in this affine matrix?

[wyli]

one possible option would be making a customized transform to assign the pixel size and SpacingD to normalize them:

import os
import numpy as np
import monai
from monai.transforms import Compose, Lambda, LoadImageD, SpacingD


class AssignPixelSize:
    def __init__(self, key, pixel_key):
        self.key = key
        self.pixel_key = pixel_key

    def __call__(self, data_dict):
        print(data_dict[self.pixel_key])
        data_dict[self.key].affine[0, 0] = data_dict[self.pixel_key][0]
        data_dict[self.key].affine[1, 1] = data_dict[self.pixel_key][1]
        return data_dict


filename = os.path.join(".", "MONAI-logo_color.png")
monai.apps.download_url("https://monai.io/assets/img/MONAI-logo_color.png", filepath=filename)
transform = Compose(
    [
        LoadImageD(keys="image", image_only=True, ensure_channel_first=True, dtype=np.uint8),
        Lambda(func=AssignPixelSize(key="image", pixel_key="pixel_size")),
        SpacingD(keys="image", pixdim=(0.1, 0.1)),
    ]
)
test_data = {"image": filename, "pixel_size": (0.18, 0.17)}
result = transform(test_data)
print(f"image data shape:{result['image'].shape}")
print(result["image"].affine)

[hermancollin]

@wyli thank you! This is exactly what I was looking for. I tested it and it works as expected.