Further generalize network models - Githubissues

DeepRegNet / DeepReg

Medical image registration using deep learning

Apache License 2.0

564 stars 76 forks source link

Further generalize network models #648

Closed mathpluscode closed 3 years ago

mathpluscode commented 3 years ago

Subject of the issue

For now, we are always concatenating the moving and fixed images together like https://github.com/DeepRegNet/DeepReg/blob/main/deepreg/model/network.py#L276

But this should not be part of the assumption. Ideally,

DDF model just take moving/fixed images to produce a DDF, then using moving/fixed labels to calculate loss if data are labeled.
DVF model just take moving/fixed images to produce a DVF, which will be integrated to get DDF, then using moving/fixed labels to calculate loss if data are labeled.
conditional model takes moving/fixed images and moving label, to predict fixed label, then using label loss.

How the DDF or DVF or predicted fixed label are calculated, this should not be part of the model.

mathpluscode commented 3 years ago

WIth the current registry system, it is not easy to change the way dealing with the input images. Therefore I'm closing this issue.