Questions about data usage in ANHIR challenge

The data is divided into training and evaluation. The training data has landmarks, while the evaluation data does not. I have the following two questions when I try to make registration based on deep learning. It would be very grateful if the organizer of the challenge could answer the questions.

  1. in one image group, such as in ‘COAD_01’, S1-HE registration has the status of evaluation, while the S3-S1 registration has the status of training. Could I use the latter in training set? I have this question, because since one group of images are all from the same sample, maybe one group images(such as ‘COAD_01’ images) should be evaluation together or be training together when training neural network.
  2. The training data in the challenge is provided with landmarks, but the evaluation data in the challenge does not have landmark files. If I use the unsupervised registration method, in other word, I do not need the landmark files, can I add all the data (with status of training and evaluation) to the neural network as training data?
    Looking forward to the replies, thanks!

Reply by Jan Kybic:

add 1)
I think, this is all explained in the data description website or in our IEEE TMI article which just appeared. Briefly yes, what is designed as training is meant to be used as training, that is how the dataset was created. There is nothing to stop you not to use same image pairs for training or evaluation but you will have less data.

add 2)
Sure, it is up to you, you can use all data for unsupervised training. However, you will then have no way of doing the evaluation.

I do not know what your status is but I recommend that such questions should be first discussed with your supervisor or knowledgeable colleague.

Thanks for your replies. It is very clear now.