Wrong intensity when RandomMotion is applied after ZNormalization

[fepegar]

To reproduce:

from torchio import Image, ImagesDataset, INTENSITY
from torchio.transforms import (
    RandomAffine,
    RandomFlip,
    RandomNoise,
    RandomBiasField,
    RandomElasticDeformation,
    RandomMotion,
    ZNormalization,
)

paths_list = [[
    Image('image', '~/Dropbox/MRI/t1.nii.gz', INTENSITY),
]]

seed = None
verbose = True

transforms = [
    ZNormalization(verbose=verbose),
    RandomMotion(proportion_to_augment=1, seed=seed, verbose=verbose),
    # RandomNoise(seed=seed, verbose=verbose),
    # RandomBiasField(seed=seed, verbose=verbose),
    # RandomFlip(axes=(0,), seed=seed, verbose=verbose),
    # RandomAffine(seed=seed, verbose=verbose),
    # RandomElasticDeformation(proportion_to_augment=0.1, seed=seed, verbose=verbose),
]

dataset = ImagesDataset(paths_list)

sample = dataset[0]
for i, transform in enumerate(transforms):
    transformed = transform(sample)
    name = transform.__class__.__name__
    path = f'/tmp/{i}_{name}.nii.gz'
    dataset.save_sample(transformed, dict(image=path))

@romainVala do you know why this could be? I'm probably missing some theory about the Fourier transform.

[fepegar]

Actually this is expected because of this line: https://github.com/fepegar/torchio/blob/master/torchio/transforms/random_motion.py#L216.

I'm not sure it's the best approach.

[romainVala]

Hi
Ok I see, MRI is a complex number but we always take the abs value.
I am not sure why this is so, but with normal acquisition we do not have negative number. So here after the ZNormalization you end up with negative number which are turn to positive after the abs operation donne after the inverse fft

You could change this line L216 to take the real part, but it is not the way the acquisition works ...

I just check to perform random motion with the same seed, and compare the 2 cases (real and abs value after inv fft) this induce changes especially in the background we you get negative value if you take the real part

So I do not thing it is a good idea to change abs to real ... (since this is not what you will get from the scanner after motion ... (the scanner is always returning the abs))

May be you just need to invert the order of the transforms ... (starting with random motion ... ?)

[romainVala]

Not realated, but I look closer to the znormalization and I do not understand the rational to compute the mean on the mask build for pixel above the mean ...

[fepegar]

I agree. What about emitting a warning if there are negative values in the input?

…

On Fri, 10 Jan 2020, 08:36 valabregue, ***@***.***> wrote: Hi Ok I see, MRI is a complex number but we always take the abs value. I am not sure why this is so, but with normal acquisition we do not have negative number. So here after the ZNormalization you end up with negative number which are turn to positive after the abs operation donne after the inverse fft You could change this line L216 to take the real part, but it is not the way the acquisition works ... I just check to perform random motion with the same seed, and compare the 2 cases (real and abs value after inv fft) this induce changes especially in the background we you get negative value if you take the real part So I do not thing it is a good idea to change abs to real ... (since this is not what you will get from the scanner after motion ... (the scanner is always returning the abs)) May be you just need to invert the order of the transforms ... (starting with random motion ... ?) — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub <#37?email_source=notifications&email_token=ADAZVVCRNXUKBJ43T5TXUXTQ5AXQ3A5CNFSM4KE2SFCKYY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGOEITDN7Y#issuecomment-572929791>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/ADAZVVB5Q5VIG5URTSNCKWTQ5AXQ3ANCNFSM4KE2SFCA> .

[romainVala]

yes good idea, since it can easily happen as you find out ...

add a warning, something like that :
Warning input volume has negative value. Using motion simulation will end up with only positive value, due to the abs value use after inverse fft