New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
min_train_masks option #15
Comments
Dear Ron, Thanks for the issue. I am personally not comfortable adding this option to the Cellpose extension, even though it is available within Cellpose... From the QuPath extension cellpose, we do not make use of the "human-in-the-loop" training options, meaning that we need dense annotations. What you are doing would create a cellpose model that would detect only mitotic cells and ignore everything else, but you might run into the issue that with very few labeled data, training will not yield appropriate results unless you add many more images... Is your goal only to find the mitotic cells? In which case a better stragegy would be as follows
I am pinging @carsen-stringer in case my suggestions would need more details or if I am wrong. If you wish to discuss this more with me and the other developers, feel free to open a topic on the image.sc forum Happy to discuss Oli |
Dear Oli, I understand your concern. The reason I wanted this option is actually for educational purposes. I would have likes to show students the difference in training a deep learning model detecting the sparsely labeled images and compare it to detecting all the cells and perform object classification afterwards. It would indeed take a lot of training images to get a good result. This would require exactly what you mention. Correcting mistakes in of the trained model, creating extra data and train again (human in the loop). Best, Ron |
Hi Ron Awesome thanks for the info! I can think of implementing it if you want. Otherwise you could simply take the command as it was printed in the log, paste into your command line and add the missing flags. The training will run, just without the fancy QA bits |
Dear Oli, Yes, please. Would be very nice if you can implement it. I also have pictures of adopicytes. These also contain just a few cells in a patch most of the time. Best, Ron |
The latest release adds the min_train_masks option Let me know if there are any issues and I will close the issue afterwards. |
Many thanks. I tested it, and everything works fine. Used min_train_mask(1) and (a bit to my suprise) mitosis are detected quite well. Trained on 17 patches in 7 images. During my course I will see how good compared with object classification. Best, Ron |
thanks for the implementation @lacan , glad it's working, this is a neat use case |
Hi,
Would it be possible to add the min_train_masks option. I try to train the detection of mitosis. But get an error like: "train images with number of masks less than min_train_masks (5)". Would be nice to set this to a lower value.
Best, Ron
The text was updated successfully, but these errors were encountered: