Small question about DeepAtlas

[mikami520]

I am now working on the DeepAtlas tutorial and have a quick question about the warp and warp_nearest in the training. Since in the tutorial, you use the alternate training mode to save memory, which means one network will be frozen while the other one is active. In the segmentation network trail, the comment (shown below) said that we could use differentiable warp() instead of warp_nearest() to achieve the joint training, but this case, the registration network is frozen, how could we do the backpropagation and gradient descent for registration network based on this learnable warp()?

In my opinion, we should not freeze the registration network this time if we want to use warp() to train the registration jointly.

Hope someone could help me and thank you in advanced!

[KumoLiu]

Hi @ebrahimebrahim, could you please help share some comments here? I think it's this tutorial you've contributed for the first time.
Thanks in advance!

[ebrahimebrahim]

In my opinion, we should not freeze the registration network this time if we want to use warp() to train the registration jointly.

Yes, that sounds right. We'd have to not freeze the registration network to do joint training.

[mikami520]

@ebrahimebrahim Thanks for your response. Therefore, if we freeze the registration network, it does not matter which warp function we use, right? But warp_nearest won't involve gradient descent(not differentiable) calculations which decrease the computational burden.

[ebrahimebrahim]

That makes sense. But I don't know for sure if using warp_nearest decreases computational burden in this case. It might, but I would think that if you are keeping the registration network frozen then it doesn't matter if you use a differentiable warp function. Not certain about this.

[mikami520]

Sounds good, thank you so much. I will have a try to test if that decreases the computational burden