RandGaussianNoise standard deviation parameter

[bakert1]

Hi. I noticed some unexpected behavior regarding the RandGaussianNoise transform (link to source).

The transform has three self-explanatory parameters prob, mean, and std.

Based on the docstring, my expectation is that std controls the spread of the Gaussian noise. In other words, if prob=1, mean=0, and std=100, then this transform should always add Gaussian noise with zero mean and a standard deviation of 100 to all image pixels/voxels.

However, this expectation isn't met, at least based on my reading of the code implementation. If prob=1, mean=0, and std=100, then this transform adds Gaussian noise with zero mean and a standard deviation randomly selected from 0 to 100. (e.g., it could with a standard deviation of 33). The standard deviation of the noise changes on every call to the transform.

Basically, I would expect the Gaussian noise to always have the same, specified standard deviation of std, especially since std is specified as a float and not a tuple (e.g., (0.0, 1.0)). I understand many other random transforms sample parameters (as is being done here with std,) but this behavior still doesn't seem intuitive to me.

Does anyone have any insight to why RandGaussianNoise is implemented in this manner? Alternatively, is anyone else surprised by this implementation?

Edit: I condensed my question for brevity. It looks like RandRicianNoise (a noise type related to Gaussian noise) has a similar implementation. Except for RandRicianNoise, there is a parameter sample_std which controls whether the standard deviation is sampled from a Uniform distribution or if it's taken as given.

[KumoLiu]

Hi @tim-the-baker, thanks for your interest here.
The RandGaussianNoise transform in MONAI is designed to introduce randomness into the model's training, aiding in better generalization and preventing overfitting. It works by adding random Gaussian noise to the image data. Rather than having a fixed noise for all augmentations, the transform adds different noise levels sampled from the specified standard deviation each time it's applied. This brings more variation and randomness, and hence robustness, to the training.

Hope it helps, thanks.

[bakert1]

Thanks for your response! That implementation makes sense to me, especially in the context of training neural network.

I wanted to use RandGaussianNoise to test the robustness of an image registration algorithm when the images are corrupted with varying levels of Gaussian noise. For this application, I want the standard deviation to be fixed and not sampled. It was easy to extend RandGaussianNoise and remove the std sampling.

I think adding the option to not sample the standard deviation to MONAI would also be useful so I went ahead and added a feature request issue that I also offered to work on.

[irecana]

Hi @tim-the-baker and @KumoLiu! Good to know that also the standard deviation is randomly picked, thank you for sharing!
I just started working in the field and I have a big doubt regarding RandGaussianNoise. What is the correct way to set the random state so that I ensure the generated noisy data are the same at each run? Currently I am setting "set_random_state(seed=0)" to the instance transform before applying it to my data, but I see some variability at each run.
I hope you can give me some hints, and thank you in advance!

[KumoLiu]

Hi @irecana, you can use set_determinism to ensure the deterministic.

MONAI/monai/utils/misc.py

Line 316 in ff43028

def set_determinism(

Here is an example:
https://github.com/Project-MONAI/tutorials/blob/main/3d_segmentation/spleen_segmentation_3d.ipynb
Thanks.

[irecana]

Thank you for your reply, @KumoLiu!
I tried that way but it was not working, not sure why. I want to specify that I call RandGaussianNoise inside a class (Transform) that I defined and that class is later called with Compose. Anyway, I found an error and using "set_random_state" on the instance transform inside the class I created is now working for me.
I have another question though: is there a way to still have some variability in the applied noise between each sample (25 samples of my validation set), but ensuring that at each run each sample has always the same noise that was applied in the previous run? Because what happens now is that each sample will have the same noise applied, that is not exactly what I want.

I cannot think of a way, rather than saving the samples, instead of running the noise transform each time, but if you have any suggestion it will be highly appreciated. Thank you again!

[KumoLiu]

Hi @irecana, if you write a custom transform yourself, you should also ensure your transform follows our convention to ensure determinism.

MONAI/monai/transforms/intensity/array.py

Line 85 in ff43028

class RandGaussianNoise(RandomizableTransform):

MONAI/monai/transforms/intensity/array.py

Line 106 in ff43028

def randomize(self, img: NdarrayOrTensor, mean: float | None = None) -> None:

For each call, the transform will generate a different parameter.
Thanks.