search

MIC-DKFZ / MedNeXt

[MICCAI 2023] MedNeXt is a fully ConvNeXt architecture for 3D medical image segmentation.
https://arxiv.org/pdf/2303.09975
Apache License 2.0
348 stars 26 forks source link

Deep supervision #1

Closed ikboljon closed 1 year ago

ikboljon commented 1 year ago

Hi.

The paper is a great read. I have a question regarding deep supervision, which is not explained in the paper, but you have that in the main figure. Did you use deep supervision for all the other models that are used for comparison? If not, how can we make sure that the performance improvement is not caused by deep supervision (which theoretically is the case)?

Thank you for the response.

Regards, Ikboljon Sobirov

saikat-roy commented 1 year ago

Hey Ikboljon,

Thanks for the question. Our main competitor during development was nnUNet - both nnUNet and MedNeXt uses deep supervision while allows for an equivalent comparison between these 2 nets.

Since a majority of the other baselines (4 out of the 6) didn't use deep supervision in their original publications - we decided to keep things uniform and not use deep supervision in all 6 model. However, you are right to point this out as this was raised in the MICCAI reviews as well - we hope to follow-up on this work and test deep-supervision to the 2 out of 6 baselines that used them in the original publication in later work.

Regards, Saikat

ikboljon commented 1 year ago

Dear Saikat,

Thank you very much for the clarification.

Regards, Ikboljon Sobirov