correction matrices from ants.motion_correction (antspy)

[garincle]

Hi,
I would to know if their is a way to convert or to get the movement correction matrices from ants.motion_correction (antspy). It seems that with the output we can only get the .mat, and the output image.
Best,
Clément

[cookpa]

To convert the format, this should do it:

tx = ants.read_transform("input.mat")
ants.write_transform(tx, "output.txt")

See also here

https://github.com/ANTsX/ANTsPy/wiki/ANTs-transform-concepts-and-file-formats

for more on how to get at the transform parameters in python.

[garincle]

Thanks but this gives 9 different parameters, I do not think they would fit with what is usually required for signal regression (often is expected 6 parameters (trans_x, trans_y, trans_z, rot_x, rot_y, rot_z)?

Working with AI, sorry if it totally stupid, I got:

def extract_motion_params_antspy(matrices_dir, output_1D):
    """
    Correctly extracts motion parameters from ANTs transformation files
    """
    mat_files = sorted(Path(matrices_dir).glob('*_0GenericAffine.mat'))
    
    if not mat_files:
        raise FileNotFoundError(f"No *_0GenericAffine.mat files found in {matrices_dir}")

    motion_params = []

    for mat in mat_files:
        # Read the transformation
        transform = ants.read_transform(str(mat))
        
        # Get the 4x4 matrix
        matrix_4x4 = np.array(transform.parameters).reshape(4, 4)
        
        # Extract the 3x3 rotation matrix
        rotation_matrix = matrix_4x4[:3, :3]
        
        # Extract translations (last column, first 3 rows)
        translations = matrix_4x4[:3, 3]
        
        # Convert rotation matrix to Euler angles
        # Robust method to avoid gimbal lock
        sy = np.sqrt(rotation_matrix[0,0]**2 + rotation_matrix[1,0]**2)
        
        if sy > 1e-6:  # No gimbal lock
            roll = np.arctan2(rotation_matrix[2,1], rotation_matrix[2,2])
            pitch = np.arctan2(-rotation_matrix[2,0], sy)
            yaw = np.arctan2(rotation_matrix[1,0], rotation_matrix[0,0])
        else:  # Gimbal lock case
            roll = np.arctan2(-rotation_matrix[1,2], rotation_matrix[1,1])
            pitch = np.arctan2(-rotation_matrix[2,0], sy)
            yaw = 0
        
        # Convert to degrees
        roll_deg = np.degrees(roll)
        pitch_deg = np.degrees(pitch) 
        yaw_deg = np.degrees(yaw)
        
        motion_params.append([roll_deg, pitch_deg, yaw_deg, 
                             translations[0], translations[1], translations[2]])

    # Save parameters
    motion_params = np.array(motion_params)
    np.savetxt(output_1D, motion_params, fmt="%.6f", delimiter=" ")
    
    print(f"[INFO] {len(motion_params)} volumes processed")
    print(f"[INFO] Parameters saved: roll pitch yaw Tx Ty Tz (deg, mm)")

What do you think?
Thank you again!

[cookpa]

At a glance I think you have the right idea, but I've not run the code.

The ordering of the parameters is described at the link above. For 3D it's the matrix (9 elements) in row order, then the translation parameters.

You can make rotation / translation matrices from known axis / angle values using ITK-SNAP, in order to test (see Tools -> Image Free Rotation)