NormalizeIntensity parameters cannot be scalars?

[agaldran]

Hello!

So, I dont think I understand correctly this transform. If I use in my code:

tr_transforms = transformsCompose([
...
transforms.NormalizeIntensityd(keys=('img', ), channel_wise=True, nonzero=True)
...
])

then it works nicely, but when I try to pass it a mean and a std with which I would like my data to by normalized (I am assuming im = (im-mean)/std, then it crashes:

m, s = torch.tensor(347.14), torch.tensor(120.35)
tr_transforms = transformsCompose([
...
transforms.NormalizeIntensityd(keys=('img', ), subtrahend=m, divisor=s, channel_wise=True, nonzero=True)
...
])

so, what should i pass to this transform instead?

Thanks!

[KumoLiu]

Hi @agaldran,
Yes, you are right, if you pass channel_wise=True, it will normalize channel-wise.

MONAI/monai/transforms/intensity/array.py

Line 894 in 6c9e49e

if self.channel_wise:

MONAI/monai/transforms/intensity/array.py

Line 907 in 6c9e49e

img = self._normalize(img, self.subtrahend, self.divisor)

MONAI/monai/transforms/intensity/array.py

Line 885 in 6c9e49e

img = (img - _sub) / _div

Hope it helps, thanks.

[agaldran]

Hey thanks for the quick answer. So it seems that with channel_wise=False things work as expected. I was thinking that with only one channel scalars would have worked the same, but that's not true.

Does the normalization expect a list with as many items as channels? Would it work if I use a list of len 1 for both the mean and the std? Thanks!

tr_transforms = transformsCompose([
...
transforms.NormalizeIntensityd(keys=('img', ), subtrahend=[mean,], divisor=[std,], channel_wise=True, nonzero=True)
...
])

[KumoLiu]

The data should be channel first array.

MONAI/monai/transforms/intensity/array.py

Line 890 in 6c9e49e

Apply the transform to `img`, assuming `img` is a channel-first array if `self.channel_wise` is True,