Is it possible to use RandShiftIntensity to apply to a specific area in an image?

[Minxiangliu]

RandShiftIntensity is one of the good data augmentation methods, now I have images with 0 values in the background, and then I only want Shift on the foreground (not the background), is this possible? Or can anyones suggest me how to do it?

Thanks in advance!

image:

[yiheng-wang-nv]

Hi @Minxiangliu , RandShiftIntensity cannot support your requirement directly. You may need to define a customized transform. Here is an example:

from monai.utils.enums import TransformBackends
from monai.transforms.transform import Transform
from monai.transforms import RandShiftIntensity
import torch

class FgShift(Transform):

    backend = [TransformBackends.TORCH]

    def __init__(self, offsets, prob = 0.1, bg_value = 0.0):
        self.shifter = RandShiftIntensity(offsets=offsets, prob=prob)
        self.bg_value = bg_value

    def __call__(self, img, offset = None):
        """
        Apply the transform to `img`.
        """
        out = torch.where(img != self.bg_value, self.shifter(img), img)

        return out

The following is a simple test code to show the result:

trans = FgShift(prob=1.0, offsets=(1, 4), bg_value=0.0)
a = torch.tensor([[-1.0, 2.5, 3.0], [1.1, 0, 0]])

for i in range(5):
    print(f"output is: {trans(a)} offset is: {trans.shifter._offset}")
    print(" ")

[Minxiangliu]

Thank you for your concise explanation, I can understand what you mean. In addition, can the method of monai.transforms.RandScaleIntensity be similar to your approach? This is because my background is 0, so multiplying by any number won't change.

[yiheng-wang-nv]

Hi @Minxiangliu , since your bg_value is 0, you can use RandScaleIntensity directly and do not need to wrap it into a new transform.