What is the method of ensembling in nnUNet? how do the 5 models (of 5 folds) vary?

[NouraAAR]

Hello,

I am trying to understand how the ensembling actually works in nnUnet, in the case of using only one configuration. What does the ensembling really do after the 5 fold cross validation?
I haven't found the answer to my question in the documentation or the paper/supplementary information.
I have used only one configuration (2D), and I have the following inquires:

1. How does each model (corresponding to each fold) vary? Does the model of each fold have a different architecture or hyperparametrs? If so, when performing final inference using all 5 folds, which architecture/hyperparameters are used?

2. I tried running inference using 1 fold (the fold that gave the best metrics), and using all folds, and the results on the test set were not identical. Therefore, I would like to know what happens during ensembling of the folds (keep in mind I'm only using 1 configuration), are the predictions of each model averaged, taking the majority vote, or the 5 architectures are somehow combined?

3. Please correct me if I'm wrong: it is my understand that the following happens during 5 folds CV using 1 configuration:

• The train set is split in 5
• 4 sets are used for training and 1 set is kept for validation, repeated 5 times for 5 folds
• This results in training of 5 models (each with a different architecture or hyperparamets? question 1)
• Inference on test set is done using the 5 folds, where the predictions of each of the 5 models on the test set is averaged to obtain the final predicition (or perhaps voted? or some other method? question 2)

Thank you in advance, answering these question will be of great help to better understand how this tool is working!

[WiDayn]

I think it averaged the output of the five models, which is mentioned in the documentation.(https://github.com/MIC-DKFZ/nnUNet/blob/master/documentation/how_to_use_nnunet.md) - Model training - Overview

"a natural way of obtaining a good model ensemble (average the output of these 5 models for prediction) to boost performance."