search

HenriquesLab / ZeroCostDL4Mic

ZeroCostDL4Mic: A Google Colab based no-cost toolbox to explore Deep-Learning in Microscopy
MIT License
550 stars 129 forks source link

Training models to identify multiple organelles with 3D UNET #142

Open Eddymorphling opened 2 years ago

Eddymorphling commented 2 years ago

Hi folks, Not sure if this is already a feature or can be incorporated into google colab. But, I was wondering if it would be possible to train and create models for multiple organelles using the ZeroCost-3D UNET google colab pipeline?

For example: can we setup a pipeline where we can create separate labels for both lipid bodies and mitochondrion on the same training 3D volume (taken with the FIB SEM maybe?) and then use the resulting UNET model to predict ER and mitochondrion on a new FIB-SEM dataset?

Eddymorphling commented 2 years ago

Or in other words, can we setup a Multi-Class-3UNET pipeline?

guijacquemet commented 2 years ago

Hi @Eddymorphling !

It would be certainly possible to create a multi-class 3D Unet. However not sure this will happen anytime soon as we do not have anyone working on this at the moment.

However, what you describe can also be achieved by training multiple single-class Unet models. One model per organelle you want to segment. Then your prediction pipeline would be to use one model after the other.

Cheers

Guillaume