nnUnet performance on BrATS2020 segmentation task

[janspiegel]

Dear all,

many thanks for nnUNET, it's been extremely helpful as a framework to get my head around brain tumour segmentation tasks - as well as what is involved in designing/running UNet architectures.

I followed the instructions provided to run the 3d full resolution Unet on the BraTS2020 data, and uploaded my predicted segmentation labels to the BrATS2020 website.

The average DICE_WT achieved was 76.53% which felt relatively low given the results reported in some Unet papers on BrATS2020 segmentation, and compared to Fabian's nnUNET paper here: https://arxiv.org/abs/2011.00848

I followed the vanilla instructions here https://github.com/MIC-DKFZ/nnUNet/blob/master/documentation/how_to_use_nnunet.md

and ran !nnUNetv2_train DATASET_NAME_OR_ID 3d_fullres 0 -device cuda --c

which gave EMA pseudo Dice: 0.8958 after 50 epochs and DICE_WT on unlabelled validation data of 0.765 .

Are these results sensible or am I missing anything obvious which means nnUNET is not performing as expected on this task?