Why do I have to do 2D DCT twice to get a sinogram?

Question

I am trying to compute 2D DCT (Discrete COsine Transform) on an tomographic image with tensorflow (in order to use it in a optimisation problem later) but doing 2D DCT once doesn't give me good results. I have to do it twice to get a sinogram. good sinogram

Radon image with one 2D dct

My code is the following

import tensorflow as tf
import skimage.io as skio
import numpy as np
import h5py 

img_data = h5py.File('LoDoPaB-CT_dataset/ground_truth_test/ground_truth_test_000.hdf5', 'r')
print(list(img_data.keys())) 
img_dset = img_data['data']
image = img_dset[20,:,:]
image = image.astype(np.float32) / image.max()

# Convert the image to a TensorFlow tensor
image_tensor = tf.constant(image)

NUM_ANGLES = 100
theta = np.linspace(5., 180., NUM_ANGLES, endpoint=False) # CHECK CHOICE OF THE ANGLES

# Reshape the image tensor to add a batch dimension
image_tensor_batched = tf.expand_dims(image_tensor, axis=0)
image_tensor_batched = tf.expand_dims(image_tensor_batched, axis=0)

def dct_2d(feature_map, norm=None ):
    X1 = tf.signal.dct(feature_map, type=2, norm=norm)
    X1_t = tf.transpose(X1, perm=[0, 1, 3, 2])
    X2 = tf.signal.dct(X1_t, type=2, norm=norm)
    X2_t = tf.transpose(X2, perm=[0, 1, 3, 2])
    return X2_t

dct_tensor = dct_2d(image_tensor_batched)

# Remove the batch dimension and convert the result to a NumPy array
dct_array = tf.squeeze(dct_tensor)

with tf.compat.v1.Session() as sess:
    array = sess.run(dct_array)

skio.imshow(radon(array, theta=theta, circle=False), cmap='gray')