Is it possible to use one network (or shared weights) instead of GX and GY?

[xiaogengen-007]

Hi,

Thank you for the nice work and for sharing the code for CycleMorph. I would like to incorporate the framework into our research on CT registration. But I have a question after reading the CycleMorph paper (CycleMorph: Cycle consistent unsupervised deformable image registration (Medical Image Analysis, Boah Kim et al.).

In CycleMorph, you trained 2 networks (GX and GY) for the forward and reverse paths, respectively. Is it possible to use only one network to do both forward and reverse paths (i.e. only switching the order of the source and the target)? In other words, can GX and GY share the same weights? If not, what is the difference between GX and GY after training? If I am understanding it correctly, there is no specific ordering between the source and the target (or X and Y).

Ideally, for a registration network, I would expect it to give both forward and reverse results (only by switching the order of the input image) if there is no specific ordering in the data.

Thank you so much in advance!

[HazelHik]

Hi,

Thank you for the nice work and for sharing the code for CycleMorph. I would like to incorporate the framework into our research on CT registration. But I have a question after reading the CycleMorph paper (CycleMorph: Cycle consistent unsupervised deformable image registration (Medical Image Analysis, Boah Kim et al.).

In CycleMorph, you trained 2 networks (GX and GY) for the forward and reverse paths, respectively. Is it possible to use only one network to do both forward and reverse paths (i.e. only switching the order of the source and the target)? In other words, can GX and GY share the same weights? If not, what is the difference between GX and GY after training? If I am understanding it correctly, there is no specific ordering between the source and the target (or X and Y).

Ideally, for a registration network, I would expect it to give both forward and reverse results (only by switching the order of the input image) if there is no specific ordering in the data.

Thank you so much in advance!

I have the same question, did you test if it's okay to train a single G network? Thanks a lot :)