Add new segmentation label into pre-trained model (wholeBrainSeg Large UNEST segmentation).

[anuragp2018]

Hello,

I want to add a new segmentation label to the WholeBrainSeg Large UNET segmentation model and retrain the model. Can you please help or guide me on how to do that?

[ericspod]

Hi @anuragp2018, I would suggest starting with the segmentation template bundle. If your model is just a UNet network then the template would require minimal modification to become your bundle. Copy the whole bundle directory, modify the configurations and documentation in README.md, provide a license, and whatever other code you need in the scripts directory. From there you can propose a PR and we'll help with anything more that needs doing.

[anuragp2018]

Actually, I don't have any model. I just want to add a new segmentation label to a pre-trained model and retrain the model.

[ericspod]

In this case you're going to want to do some form of transfer learning. One way is to first load the model and then replace the last block with one having a different number of output channels, then either fine tune just that block or the whole network on your dataset. This tutorial may help with that, otherwise @tangy5 or @yiheng-wang-nv may have some insights as well.