Creating A Custom Annotation Model From A Custom Segmentation Model

Hi,

This is a terrific tool, congrats to the NVIDIA team behind this!

Presumably AI assisted annotation allows creation of new ground truth (eg for transfer learning) where the existing pre-trained segmentation model only approximates the desired segmentation, which can then be used to train a more accurate model.

After training a new segmentation model with new ground truth made with AI assisted annotations, how might one convert this new segmentation model into a new annotation model for another round of transfer/active learning?

In brief, how is the model annotation_ct_liver different from segmentation_ct_liver? How could I take a custom segmentation model segmentation_ct_custom_liver and create a custom annotation model annotation_ct_custom_liver?

Thanks for your help!

-Brett

Hi

Thanks for your interest in clara SDK and your nice complements. I will make sure to pass it on to our engineering team.

We briefly touch on the difference between annotation models and segmentation in our webinar https://info.nvidia.com/accelerate-discoveries-with-the-nvidia-clara-ai-toolkit-reg-page.html

Basically segmentation require 1 channel (image) while annotations requires 2 channels (image and user hints). However, using just the segmentation you can train an annotation model and use our transformations to simulate the user inputs (see AddExtremePointsChannel transformation) in any of our annotations models.

For a full cycle you should be training 2 models as you get more data a segmentation model to do predictions and an annotation model to help you generate more data.

Hope that helps

Thank you, this makes sense now!