Difference between Conference and Journal Submissions (BraTS)

[stalhabukhari]

Hi @FabianIsensee

I hope you're doing great. I read your conference paper on Brain Tumor Segmentation, and recently read the journal version that discusses other applications as well. I wish to inquire specifically about Brain Tumor Segmentation.

I noticed that the nnUNet reported in the journal version is slightly different from the one reported in the BraTS 2018 challenge. Specifically, the differences included the use of Leaky-ReLUs instead of ReLUs, use of auxiliary loss functions for deep supervision, the absence of any mention of region-based prediction (similar to the cascade of Wang et al. (2017)), and the use of SGD + Nesterov Momentum (with a per-epoch decaying schedule for learning rate) instead of Adam optimizer with constant learning rate. Kindly correct me if I missed something.

Are the results/plots for Brain Tumor Segmentation shown in the journal version different from the conference one? If so, then which version would you prefer specifically for Brain Tumor Segmentation? Was any comparison done in this regard?

Best Regards,
Talha

[FabianIsensee]

Hi,
the arxiv paper you are referencing is NOT nnU-Net. It is a separate project that happens to have a badly chosen title ;-)
For brain tumor segmentation on the BraTS datset I would stick to our BraTS 2020 implementation which is based on nnU-Net: https://arxiv.org/abs/2011.00848. Code for that is in the nnunet repository. If you are using a different dataset (and a non-region based evaluation) then using the standard nnU-Net is best
Best,
Fabian

EDIT: I changed the word 'jounal' -> 'arxiv' in the first sentence. That was a mistake. Now everything is correct: The "No New Net" arxiv paper is NOT nnU-Net! The Nature Methods journal paper IS nnU-Net :-)

[stalhabukhari]

the journal paper you are referencing is NOT nnU-Net. It is a separate project that happens to have a badly chosen title ;-)

Oh! Then how should I interpret the results for brain tumor segmentation in the journal paper? :D

For brain tumor segmentation on the BraTS datset I would stick to our BraTS 2020 implementation which is based on nnU-Net: https://arxiv.org/abs/2011.00848.

Thank you for the clarification. I'll go through it in detail and will let you know then.

Code for that is in the nnunet repository.

But the journal paper points to this repository under the section Code availability. Should I refer to different repositories for the papers I mentioned?

If you are using a different dataset (and a non-region based evaluation) then using the standard nnU-Net is best

Alright!

[FabianIsensee]

The Nature Methods paper only uses Task01 from the MSD for brain tumor segmentation. That is all the old (and not very well annotated) BraTS data. I would recommend to stick to our 2020 arxiv paper for BraTS-like brain tumor segmentation and for all other brain tumor segmentation just apply the nnU-Net from the journal paper.

This repository is the code for the journal paper. Within this repository is a small section that covers the BraTS 2020 - specific adaptations that we made.
You can find them here: https://github.com/MIC-DKFZ/nnUNet/tree/master/nnunet/training/network_training/competitions_with_custom_Trainers

And here: https://github.com/MIC-DKFZ/nnUNet/blob/master/nnunet/dataset_conversion/Task082_BraTS_2020.py

[FabianIsensee]

Ah shoot. I meant to say that the BraTS 2018 arxiv paper "No New Net" is NOT nnU-Net! haha. That was something else that was just named badly.

The journal paper of course is. This repository is for the journal paper.

[stalhabukhari]

Ah shoot. I meant to say that the BraTS 2018 arxiv paper "No New Net" is NOT nnU-Net! haha. That was something else that was just named badly.

I thought so! Okay, thank you for your time!