External data: use of predictions from other models

The rules state "The data used to train algorithms are restricted to the data provided by the challenge."

I wanted to clarify couple of things: 1) Does this mean that predictions of any other model (e.g. segmentation masks or registrations from other models) cannot be used to pre-process the data or train the translation models? 2) If the use of predictions of other models is dis-allowed, then does that apply to validation phase as well?

Hi Saad,

Thank you for reaching out. Interesting case. With the rule we mean that the data used to perform synthesis need to be restricted to the provided ones. In your case, it seems that you wish to use other model to pre-process the data. This would be ok, in case you are using the model to generate some masks or pre-register the data. If the pre-trained model is part of the training, this would not be allowed. Why? The pre-trained model may introduce an optimistic bias in the results you may obtain, and we wish to run a fair competition among the participants.

The restriction is extended to the whole challenge, so validation as well.

A reminder to overcome future issue: the run time per each case should be 15 minutes. Using an additional model to pre-process the data may lengthen the process, so make sure you are within the allowed time. Also, you will need to dockerize any additional model in an algorithm: it is surely feasible, but good to consider also this point.

I hope I have addressed your questions. If something may still be unclear, please, do not hesitate to answer here.

Cheers,

Matteo