Based on this wrapper also create an easy way to evaluate the results on a test set with labels.
Provide and clean up scripts for the training.
- The exact training scripts used for the paper have been moved into their own folder
training_scripts/oai_paper_pipeline/
(done)
Document.
For the OAI pipeline we need to support image to atlas-image registration. This will be different than image-to-image registration as the atlas will have a slightly different appearance (it is in the repo). Not sure how much difference that will make. It might require a different similarity measure (maybe NCC) and/or might require a different kind of training (as not all the image-pair combinations will be useful, instead it will be atlas-to-image and image-to-atlas).
Similarity measure: If another similarity than SSD is required (e.g., NCC or something more sophisticated) is the ICON approach still expected to work?
- NCC is implemented in the library and works, although with lower final accuracy
Training strategy: Since the goal is image-to-atlas registration there will in principle be fewer training pairs (as not all pairwise image combinations are feasible). Will this likely create issues for the approach? I suspect one could always train an image-to-image approach first and then fine-tune for image-to-atlas. Another (maybe more complex alternative) would be to use a similarity measure as in Zhipeng's paper (attached).
For the registration interface it might make sense to make it a little more general. E.g., where it would also allow image-segmentations as inputs (if desired).