ITKReader doesn't return numpy array

[Nic-Ma]

Describe the bug
Currently, we use the itk.array_view_from_image() to get data array:
https://github.com/Project-MONAI/MONAI/blob/dev/monai/data/image_reader.py#L316
But actually, it's not numpy array, so following logic which depends on numpy.ndarray type will fail.

loader = LoadImage()
img, meta = loader(dicom_path)
print(type(img))

Get the output:

<class 'itk.itkPyBufferPython.NDArrayITKBase'>

[wyli]

NDArrayITKBase is a subclass of ndarray, shouldn't we enhance/fix the relevant failing logic instead of converting them to ndarray, is this a duplicate ticket of #2885?

[Nic-Ma]

Hi @wyli ,

Thanks for your suggestion, makes sense.
I updated the PR to update the relevant failing logic instead: #2900.
Could you please help review it?

Thanks.

[Nic-Ma]

Hi @wyli ,

After some testing, I feel it's better to also update ITK reader to use np.asarray(), because some operations of NDArrayITKBase is different from ndarray, very easy to cause unknown errors. For example:

print(type(img))
ret = np.mean(img)
print(isinstance(ret, np.ndarray))
print(ret)

img1 = np.array([1, 2, 3])
print(type(img1))
ret1 = np.mean(img1)
print(isinstance(ret1, np.ndarray))
print(ret1)

The output is:

<class 'itk.itkPyBufferPython.NDArrayITKBase'>
True
-3.4086406e-07
<class 'numpy.ndarray'>
False
2.0

The check for isinstance(mean, np.ndarray) will fail here.

What do you think?

Thanks.

[wyli]

I think the problem is still that monai's current type conversion is not robust https://github.com/Project-MONAI/MONAI/blob/dev/monai/utils/type_conversion.py#L155, we should try to enhance it, instead of removing all the potential additional features brought by NDArrayITKBase.

[Nic-Ma]

Hi @wyli ,

Not only the type conversion, the above np.mean() example comes from the NormalizeIntensity transform:
https://github.com/Project-MONAI/MONAI/blob/dev/monai/transforms/intensity/array.py#L593
It will cause error in the next line:

if isinstance(_sub, (torch.Tensor, np.ndarray)):
    _sub, *_ = convert_to_dst_type(_sub, img)
    _sub = _sub[slices]

Thanks.

[wyli]

After some testing, I feel it's better to also update ITK reader to use np.asarray(), because some operations of NDArrayITKBase is different from ndarray, very easy to cause unknown errors. For example:

print(type(img))
ret = np.mean(img)
print(isinstance(ret, np.ndarray))
print(ret)

img1 = np.array([1, 2, 3])
print(type(img1))
ret1 = np.mean(img1)
print(isinstance(ret1, np.ndarray))
print(ret1)

The output is:

<class 'itk.itkPyBufferPython.NDArrayITKBase'>
True
-3.4086406e-07
<class 'numpy.ndarray'>
False
2.0

The check for isinstance(mean, np.ndarray) will fail here.

What do you think?

Thanks.

@thewtex might be interested in this usability tests of NDArrayITKBase

[wyli]

Closing this via #2900 for now, please feel free to reopen and discuss further.