Cloning, Editing and Adding Deep Learning Models
Note: Segmentation using Deep Learning requires the Deep Learning extension to the 2D Automated Analysis module. The Image-Pro Neural Engine must be installed. Installing the Image-Pro Neural Engine
Clone or Add New
When cloning a model you have the option to copy the current model, along with all of the existing training data. Whether to train a new or cloned model is dependent on numerous factors and there isn't a single correct strategy.
Studies indicate that extending an existing Cellpose model with additional data on new types of objects can require significantly fewer training objects than training objects from scratch (Pachitariu and Stringer 2022). Creating a new untrained model is likely to be necessary only if your objects differ markedly in appearance or scale from the objects on which an existing model was trained.
In contrast, the authors of StarDist, made the observation that "training (from scratch) using a combination of a small custom dataset together with an existing bigger (and somewhat similar) dataset can lead to better results than just training with the custom data." (https://stardist.net/docs/faq.html)
UNET was designed to work with very few training images and yield precise segmentation (Ronneberger, Fischer and Brox 2015) so cloning existing models is seldom required for this architecture.
You may also want to try using exiting training data with a different deep learning architecture. To do this while keeping your original model intact, clone the model (including the training data), then Edit the newly cloned model. You can change the architecture during the Editing step.
Cloning Models
All models, including Image-Pro AI models can be cloned. Cloned models must be given a new name. Cloned models can be both edited and extended through additional training.
Editing Models.
Editing a model allows you to make a range of changes to a model. These range from trivial changes, such as changing the name of the model, to radical changes such as selecting a different segmentation method, a different architecture or changing tile size.
A warning dialog will be displayed when your changes require that a new untrained model is created. When making radical changes such as changing model architecture cloning the model, and making edits on the cloned model is recommended.
Note: If your edits are not compatible with the model file that you are cloning, a dialog with the text 'New Parameters are inconsistent with the existing trained model file or source images already added to the training. Click OK to keep the changes and reset the trained model file'. Clicking OK will generate a new, untrained model file while keeping the training images.
Creating New Models
You can create new models that are untrained and that have no existing training data.
Bibliography
https://stardist.net/docs/faq.html
Pachitariu, M., Stringer, C. (2022) Cellpose 2.0: how to train your own model. Nat Methods 19, 1634–1641. https://doi.org/10.1038/s41592-022-01663-4
Ronneberger, O., Fischer, P., Brox T.(2015) U-Net: Convolutional Networks for Biomedical Image Segmentation.
Medical Image Computing and Computer-Assisted Intervention (MICCAI), Springer, LNCS, Vol.9351: 234--241, 2015, available at arXiv:1505.04597 [cs.CV]