Empty Segmentations layer with no predictions

[mikhaiilenko]

Dear all,

When testing your plugin, I encounter an issue at the segmentation stage : after feature extraction, model training and adding a new segmentation layer with predictions, an empty Segmentations layer is added (with no predictions, see video below). What could be the problem ?

In my case, 439_stitched_patch is an image with vessels, and 439_binary_patch is the ground truth layer. No matter what model parameters I choose or what patches I try, I get the same result. The only thing is that my files are in .npy format. Could it be that the plugin only supports .tiff, or am I doing something wrong (though it’s hard to make a mistake here) ?

Thanks in advance for your response!

Screencast.from.12-03-2025.16.13.26.webm

[mese79]

Hi,
Thanks for trying our plugin!

The problem is you didn't add any labels. In order to segment your images, you need to add a label layer and start adding labels using the brush tool as briefly mentioned here. You need to add labels at least for two classes: 1-Background and 2-Objects you're trying to segment.
The documentation is not yet complete, and we are working on that (apologies).

However, if you already have the ground truth, then perhaps you don't need to use this plugin to train an RF model. This plugin is made for reducing the manual work for creating the ground truth annotations. If you have enough ground truth annotations, then maybe training a deep learning model is more suitable for your segmentation task, but I'm not sure.

[jdeschamps]

Hi @mikhaiilenko !

Thanks for testing out FF! Could you describe briefly what you are trying to do with your images? Do you want to segment all the fluorescence in the images? We see that you have incomplete labels already, are they coming from a different algorithm?

It may be easier to start from no labels, and use the brush to add them bit by bit (see what they look like here.

Hi, Thanks for trying our plugin!

The problem is you didn't add any labels. In order to segment your images, you need to add a label layer and start adding labels using the brush tool as briefly mentioned here. You need to add labels at least for two classes: 1-Background and 2-Objects you're trying to segment. The documentation is not yet complete, and we are working on that (apologies).

As @mese79 said, all the labels in your label layer (439_binary patch) are class 1 (blue), but FF only considers foreground the other classes ( >1). Unfortunately, that's a choice to maintain the transparency of the background segmentation (which napari does on label 1).

Two things we could do better @mese79:

• We should add details about the label values to the documentation, at least make it a lot more obvious. Maybe also add a tip on how to convert class 1 into class 2.
• We should have a message pop up, or a better error message when there is only background classes.

However, if you already have the ground truth, then perhaps you don't need to use this plugin to train an RF model. This plugin is made for reducing the manual work for creating the ground truth annotations. If you have enough ground truth annotations, then maybe training a deep learning model is more suitable for your segmentation task, but I'm not sure.

Once trained it could be used to segment more images, for which there are no labels yet!

[mikhaiilenko]

@jdeschamps @mese79 Hi all !

Now everything is crystal clear to me. I apologize for my mistake, I initially thought that multiple classes would mean different types of cells that might need to be annotated. Now, after annotating both the background and the cells, segmentation has started working. Thank you very much for your explanation !

As a user, I would like to highlight that your documentation is consistent and well-structured, and in my opinion, what could be improved is adding a bit more detailed information on annotation (I initially wanted to use a binary mask generated by our previous model in the project). Additionally it might be helpful to include more details about the model sizes for feature extraction (such as the number of parameters etc) and about post-processing parameters (how to find a good fit, for example). Everything else seems very clear to me !

P.S. By the way, when I receive a notification window saying “Prediction is done!”, clicking the close (X) button not only dismisses the notification but also closes Napari entirely. This might be something you could adjust to prevent accidental layers loss ?

[mese79]

Absolutely, I need to add more documents and a tutorial on how to use the plugin.

P.S. By the way, when I receive a notification window saying “Prediction is done!”, clicking the close (X) button not only dismisses the notification but also closes Napari entirely. This might be something you could adjust to prevent accidental layers loss ?

it's a bit weird! I'm sure I keep closing the annoying pop-up, but nothing terrible has happened after that (yet)! 😁
I will test it more, maybe it's related to OS or hardware...

[jdeschamps]

Thanks a lot for the kind words and the pointers!

What OS are you using?