Calculate loss using ddf in 3D image registration without labels

[nehagoyal1994]

Hi All,

I am experimenting 3D image registration without the help of labels. To calculate the loss, I am currently using MSELoss().
I want to use the ddf in the loass calculation after each epoch; for example :

                                loss   =   image_loss(pred_image, fixed_image)   * \                                                           
                                                        regularization(ddf)                        

where               image_loss = MSELoss()  and  regularization = BendingEnergyLoss()   

Please See: https://deepreg.readthedocs.io/en/latest/tutorial/registration.html#unsupervised

So far the loss is not exponentially decreasing for 200 epochs.
Could you please guide me with loss calculation?

Thanks in advance

[Nic-Ma]

Hi @kate-sann5100 @wyli ,

Could you please help share some comments here as you are experts for registration?

Thanks in advance.

[wyli]

the image_loss * regularization looks suspicous to me, perhaps that should be image_loss + lambda * regularization?

[nehagoyal1994]

I already tried with lambda ( different values ) but each time the loss was increasing exponentially rather than decreasing