Add dice loss for training or instance optimization

[dyollb]

Would it not help in some cases to add (mean) dice loss for some datasets?

Also, would it make sense to use dice loss (when corresponding labelfields are available) in the instance optimization?

[clarkbab]

Agreed, this would be a great addition to this tool as many papers show registration improvements when segmentation labels or keypoints/landmarks are provided during training (e.g. Voxelmorph paper - https://arxiv.org/abs/1809.05231). Sounds like a lot of work to retrain the model in a supervised way - and the training data may not contain any/enough of these labels. I see the instance optimisation idea as a great alternative and much simpler to implement. It would be nice to have dice (for labels) or target registration error (for keypoints) options also.