lassoan / SlicerMONAIAuto3DSeg

Extension for 3D Slicer for running MONAI Auto3DSeg models
MIT License
66 stars 10 forks source link

Upload more quick models #23

Closed diazandr3s closed 9 months ago

diazandr3s commented 9 months ago

Body - quick Mediastinal anatomy - quick

lassoan commented 9 months ago

Thank you, I've tested the body segmentation and it works well on CTs!

What would we need to do if we wanted to improve its performance for CBCTs? Can this model be used as a starting point and be further trained with additional data? Or we would need to add CBCT images to the original training data set and train again from scratch? Maybe describing the potential solutions here would not be the best place (other people would not find it), but if you create a tutorial about how to train Auto3DSeg then you could add a section there about refining an existing segmentation model with additional training data.

diazandr3s commented 9 months ago

Thanks for your feedback, @lassoan.

What would we need to do if we wanted to improve its performance for CBCTs?

I'd suggest adding CBCTs to the training set. Do we have a public dataset for this task?

Can this model be used as a starting point and be further trained with additional data? Or we would need to add CBCT images to the original training data set and train again from scratch?

If it is for the same segments, yes! This model can be used as a starting point. Otherwise, we'll have to train from scratch.

Maybe describing the potential solutions here would not be the best place (other people would not find it), but if you create a tutorial about how to train Auto3DSeg then you could add a section there about refining an existing segmentation model with additional training data.

There are already some tutorials on how to train the Auto3DSeg on several tasks (i.e. btcv, kits23, msd, etc): https://github.com/Project-MONAI/tutorials/tree/main/auto3dseg/tasks

But I agree. This is something we have to improve. I'm thinking of adding another tutorial that focuses on this module.

I hope this makes sense,