Impact

More precisely, we use batches of size 2. Each batch contains one patch with the foreground oversampled. Furthermore, we split each silo's data into training and validation data with 80% and 20% split, respectively. All this pre-processing and patching is done using the nnU-Net library [IJK+21]. Loss function We use the same loss function as proposed by nnU-Net [IJK+21] for the KiTS19 dataset which is based on DICE [Dic45] and on the Cross Entropy loss.