Tutorials/3d_registration /paired_lung_ct.ipynb: SSIM results of the predicted image are not as expected

[lvora]

I have executed the example at https://github.com/Project-MONAI/tutorials/blob/main/3d_registration/paired_lung_ct.ipynb
I did not get the registration results expected. Is this the best that the network can perform.? Am I missing something here in terms of configuration?

Please see details:

I ran the example and the section on inference using the pretrained model/weights (did not train from scratch).

I did get similar results as on the tutorial. However, I tried to obtain an SSIM value on the predicted images (registered, output of the model) as compared to the fixed images. I got an SSIM that was even lower than the SSIM of the moving images (unregistered) as compared to the fixed images.

Here is the code for SSIM that I ran at the end of the tutorial:

from skimage.metrics import structural_similarity as ssim

ssim_value, ssim_image = ssim(val_fixed_image, val_pred_image, data_range=val_fixed_image.max() - val_fixed_image.min(), full=True)

print("SSIM between predicted and fixed:", ssim_value)

ssim_value2, ssim_image = ssim(val_fixed_image, val_moving_image, data_range=val_fixed_image.max() - val_fixed_image.min(), full=True)
print("SSIM between moving and fixed:", ssim_value2)

SSIM between predicted and fixed: 0.27654008252080553
SSIM between moving and fixed: 0.4181062033698231

The SSIM between moving and fixed is 0.41. I would expect that after registration the SSIM improves. However the SSIM is much lower.

Is this the best that the network can do? Am I missing something here?

In order to check how this compares to classical registration techniques, I did run a classical technique using a third party tool and achieved an SSIM of 0.63.
The first screenshot shows the results of MONAI registration and the second shows the results from the classical technique. As you can see classical technique seems to perform much better.

[aylward]

You may want to try UniGradIcon:
https://github.com/uncbiag/uniGradICON

The good news is that there is already work to get it ported to MONAI! #8236