Network structure problem about mambamorph

[Guo-Stone]

Yeah, the training approach is NOT unsupervised. To be accurate, MambaMorph is trained in a weak-supervised manner instead of semi-supervised. The definition of weak-supervision is that we don't supervise the model's output (i.e. deformation field in this context) directly, but supervise the label with dice loss.

[Dummy-design]

Thank you for your reply. I noticed that the loss fluctuates significantly while reproducing your work. Is this a normal phenomenon?

[Guo-Stone]

If I were you, I would plot the loss of several epochs. Only that I can see the trend clearly.