ITK Euler3D rotation matrix not supported

[tomfrankkirk]

I've generated a rigid transform using ANTs AI. When I try to load the .mat file using nitransforms, it throws an "unsupported transform" error during the Matlab dict load function linked below.

nitransforms/nitransforms/io/itk.py

Line 141 in 9486757

affine = mdict.get(

Doing a bit of digging, my transform has 6 parameters vs the usual 12. When I load it using scipy.io.loadmat, the dict has two keys, "Euler3DTransform_double_3_3" and "fixed". A typical ANTs affine has "AffineTransform_double_3_3" and "fixed". I've given the details of each transform below.

According to chatgpt the Euler3D params are expressed as [rotx, roty, rotz, tx, ty, tz]. Its proposed the following code to convert the Euler3D into a full Affine transformation and it seems to work on my simple test cases.

I'd be happy to incorporate this into the existing from_matlab_dict function but am not expert enough in this area to fully test it or consider where it may break. Does anyone have any comment to make? Is there a reason why Euler3D isn't supported; does this solution look sensible?

Euler3D rigid transform:

{'Euler3DTransform_double_3_3': array([[ 1.34468616e-04],
        [ 1.03066786e-02],
        [ 5.49512615e-03],
        [-1.88564745e+00],
        [-1.93042799e+00],
        [ 4.81095387e+00]]),
 'fixed': array([[-3.4070487 ],
        [33.00756454],
        [-3.60144258],
        [ 0.        ]])}

Affine transform:

{'AffineTransform_double_3_3': array([[ 1.03404642],
        [-0.04273591],
        [ 0.01717928],
        [ 0.03782461],
        [ 0.99473991],
        [-0.01715351],
        [-0.00832061],
        [ 0.00896228],
        [ 0.92373895],
        [-3.23504401],
        [-1.28125352],
        [ 4.6121046 ]]),
 'fixed': array([[-3.4070487 ],
        [33.00756454],
        [-3.60144258]])}

Code to convert Euler3D into standard affine matrix

    # 1. Parse Euler angles and translation
    # input params: [angX, angY, angZ, tx, ty, tz]
    orig_params = data["Euler3DTransform_double_3_3"].flatten()
    angles = orig_params[:3]
    translation = orig_params[3:]

    # 2. Parse Fixed Parameters (Center of Rotation)
    # Euler3D has 4 fixed params; generic Affine has 3.
    orig_fixed = data["fixed"].flatten()
    center = orig_fixed[:3]

    # 3. Compute 3x3 Rotation Matrix
    # ITK Euler3D uses 'xyz' sequence
    rot_matrix = R.from_euler("xyz", angles).as_matrix()

    # 4. Flatten the matrix and append translation
    # Flattening must be row-major (m11, m12, m13...)
    new_params = np.concatenate([rot_matrix.flatten(), translation])

    # 5. Prepare data for .mat file
    # We use 'double' precision as ANTs expects this
    new_data = {
        "AffineTransform_double_3_3": new_params.reshape(-1, 1).astype(np.float64),
        "fixed": center.reshape(-1, 1).astype(np.float64),
    }

    # 6. Save
    savemat(output_mat, new_data)
    print(f"Successfully converted to generic affine: {output_mat}")

[mgxd]

AFAIK, it only hasn't been implemented because we haven't run into many Euler3DTransforms. Given that it's just an optimized AffineTransform representation, seems reasonable to add support for ingestion - please feel free to open a PR if you'd like!

[effigies]

Do we definitely want to transform to affines? I would suggest keeping Euler3DTransforms as-is unless specifically requested to transform.

Any implementation should come with tests to verify that we are matching ITK's behavior.

[tomfrankkirk]

@mgxd I've changed my workflow in the end so this specific issue doesn't arise for me any more. As such I won't open a PR for it now (sorry about that), but perhaps others will come across this thread in the future.

@effigies I was transforming to affine so I could use other tools that expect a generic affine. That seemed like the route with minimal knock-on changes.