training swin UNETR with BraTS_2021 dataset

[JuanDuranMcgill]

Hello!

What an amazing github, I can tell that a lot effort has been put into it.

I would like to train the swin UNETR model https://arxiv.org/pdf/2201.01266.pdf from scratch, with the dataset.

I understand how to format the dataset so that it can work with uuNet, but I am not quite sure how to select swin UNETR and train from scratch. The information in the readme file only explains how to train a pre-trained model, with existing weights.

Could you please give me a hand on how to solve this?

Thank you very much for your amazing work,

Juan

[rixez]

Hello @JuanDuranMcgill,

I appreciate your interest in this repo.

About the changes that you'd like to implement, you can do that by making a new trainer function with the SwinUNETR network.

I just pushed a new commit showing how to do that using the SwinUNETR model from the Monai library. You can train using the same command as before, but changing the trainer to nnUNetTrainerV2BraTSRegions_SwinUNETR.

The code is for demonstration purposes, so I did not implement deep supervision in the network. You can check this file if you want to implement it: https://github.com/rixez/Brats21_KAIST_MRI_Lab/blob/0a5397607fcd82ccf74093dd2d79407f68009e32/nnunet/network_architecture/generic_UNet.py#L171

I hope that this is helpful. Let me know if you have any further question.