Project-MONAI / MONAI

AI Toolkit for Healthcare Imaging

Home Page:https://monai.io/

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

Add MedNeXt model architectures within MONAI

surajpaib opened this issue · comments

Is your feature request related to a problem? Please describe.
There seems to be some demand for ConvNeXt based architectures #4798 implemented in MONAI. Having a ConvNeXt-based segmentation model could be quite useful as it seems to perform very convincingly, based on results from this paper https://arxiv.org/abs/2404.09556

Describe the solution you'd like
Provide an implementation of the MedNeXt model architecture in MONAI

Additional context
I've implemented and benchmarked an implementation of MedNeXt, heavily based on code (https://github.com/MIC-DKFZ/MedNeXt) from the original authors. This implementation has support for dynamic configuration of blocks and follows an API very similar to SegResNetDS and was benchmarked through MONAI bundle config.

Happy to contribute this to MONAI (with the proper source attribution) if there is interest.

Hi @surajpaib, welcome to the contribution, and if you like you can also add a bundle with your benchmark result into the model zoo repo and tutorial in the tutorial repo! Thanks!

Hi @KumoLiu, great!

I will make a PR first into the core and then I can probably spend some time later making it a zoo supported bundle, right now I am using lighter with the MONAI bundle config structure and it might need some tweaks.

Hello @surajpaib,
I'm also interested by implementing MedNeXt in the MONAI core library as a first participation in an open-source project.
So I just wonder where you are with it so far and if you need help for integration.

Hi @rcremese

I've got the mednext code repurposed from the original repo to a very MONAI SegResNet like interface. I'll need to make the push to my MONAI fork. I'm currently busy on a work trip. I'll hopefully push it over the next week. It would be great to have some help with testing and simple performance benchmarks if that interests you