How to make Custom transformation on monai in a simple way

[Maryam-Mostafa]

Hi,
I want to make a custom function for pixelSpacing that takes the a parameter and I want to apply it to the key, and image. I've seen this notebook https://github.com/Project-MONAI/MONAIBootcamp2021/blob/main/day1/1.%20Getting%20Started%20with%20MONAI.ipynb , but couldn't know what to do in my case.
I don't know if it can be done using lambda ? or if it can be done by making a subclass that inherit maptransform but inside it I'm confused of what should be done here?

this is my funtion :

def pixelspacing(img,pxlsp,target_space=[0.4,0.4,0.4]):
   orig_shape= img.shape
   target_shape = [int(np.round(orig_shape[0] *  (target_space[0]/pxlsp[0] ))),
                   int(np.round(orig_shape[1] *  (target_space[1]/pxlsp[1] ))),
                   int(np.round(orig_shape[2] *  (target_space[2]/pxlsp[2] )))]

   resize_factor= np.array(target_shape)/np.array(orig_shape)
   img = ndimage.zoom(img,resize_factor, order=1)
   return img

and I want to call it like this in the compose :
Pixelspacing(("img","seg"),pxlsp) #pxlsp is the the parameter I want to send to my function

Thanks in advanced.

[wyli]

good question, one way to do it is to use the adaptor API in monai.transforms

MONAI/monai/transforms/adaptors.py

Line 13 in 30d592c

How to use the adaptor function

import numpy as np
from scipy import ndimage

from monai.transforms import Compose, LoadImageD, adaptor


def PixelSpacing(pxlsp, target_space=[0.4, 0.4, 0.4]):
    def pixelspacing(img, pxlsp=pxlsp, target_space=target_space):
        orig_shape = img.shape
        target_shape = [
            int(np.round(orig_shape[0] * (target_space[0] / pxlsp[0]))),
            int(np.round(orig_shape[1] * (target_space[1] / pxlsp[1]))),
            int(np.round(orig_shape[2] * (target_space[2] / pxlsp[2]))),
        ]

        resize_factor = np.array(target_shape) / np.array(orig_shape)
        img = ndimage.zoom(img, resize_factor, order=1)
        return img

    return pixelspacing


xform = Compose(
    [
        LoadImageD("my_img_key", image_only=True),
        adaptor(PixelSpacing([1.0, 1.0, 1.0]), "output_img_key", {"my_img_key": "img"}),
    ]
)
output = xform({"my_img_key": "./avg152T1_RL_nifti.nii.gz"})
print(output.keys())
print(output["output_img_key"].shape)

[Maryam-Mostafa]

yeah I've already changed this but still getting the same error.
adaptor(PixelSpacing(pxlsp),"img", {"img": "img"})
and I also tried this :
adaptor(PixelSpacing(pxlsp), ['img'])
and the same error appears.

[wyli]

ok, so I'm not sure how train_dataset is created, will need more details to reproduce the error you provided... the error broadcast together with shapes (3,) (4,) looks like it's some shape mismatching for example with or without the channel dimension.

[Maryam-Mostafa]

I found the error I was applying AddChannelD before the adaptor and that's why the shape mismatch error occurred.
I just have one last question, at first I wanted to use the Spacing method in transforms for making all the images with same pixel size, but I didn't understand how it works. Is there a difference between it and my own implementation could you tell me what is it?
My method is taking the original pixel spacing and the target one then I compute a ratio between them to get a factor that changes the dimensions of the image to reach the target pixel spacing.
Also my target pixel spacing must be in range [0.1,0.5] but the example in the documentation use big values like 5 so I was confusing here.

Thanks for your consideration.

[wyli]

looks like the resize factor should be resize_factor = np.array(orig_shape) / np.array(target_shape)? still it won't always compute the correct output spatial shape. I'd recommend you use or look into the monai implementation...

[Maryam-Mostafa]

Thank you for your generous help and replies 🌹 🙏🏻