The rotation given by Random Affine not accurate

[zhilothebest]

Is there an existing issue for this?

• I have searched the existing issues

Bug summary

When I ask torchio to run a random affine with rotations between -5 to +5 degrees, I get a much wider degree rotation, probably something in the ~20+ degree range, and it appears more like a skew than a true rotation

Code for reproduction

img = [ CT scan] #in tio.Subject

transform = tio.RandomAffine(scales=(1,1,1,1,1,1),
                                 degrees=(0,0,0,0,-5,5), 
                                 center = 'image',
                                 image_interpolation='bspline',
                                 label_interpolation = 'bspline',
                                 default_pad_value = 'mean',
                                p=1)

formed = transform(img)
formed = np.moveaxis(formed.image.numpy(), [3,0],[0,3])

imshow(formed[slice])

Actual outcome

original image

transformed image

Pictures to show the outcome.

Error messages

No response

Expected outcome

Ideally, if it was truly rotations around the Z axis's image center, the circle should maintain its shape as a circle and not be deformed.

System info

No response

[romainVala]

Hello

that sound weird, I do not reproduce with for instance with
suj =tio.datasets.mni.Colin27()

I would say it comes from your image (can you share it ?)
may be the z axis is not was you think

[zhilothebest]

Sadly, I am unable to share the full data file due to agreements between the owner of the dataset and I. But would love help in how to implement in any other way. Ive tried isolating with x and y axis but , y axis rotates along the j index and x axis rotates along the I index.

[romainVala]

what is the shape of the data, can you report a print of the affine

[zhilothebest]

the shapes have extra axis's for an external slicer I am using instead of matplotlib.

[zhilothebest]

The affines were in the incorrect order. That fixed it.