How to Optimize StarDist Segmentation

Dear Oliver,

thank you for the fast answer! Actually I try the pre-trained models in Fiji, I also try different parameters but the prediction is still bad.

Can I ask you the meaning of some of the parameters? In particular in advanced options

-number of tiles

-boundary exclusion

  • verbose

  • show CNN progress

-show CNN outputs ( Is CNN Convolutional Neural Networks?)

maybe if I change also these parameters I can hope to improve the prediction.

in the following link, I put a screenshot to show you the problems that I found. For example, often 2 cells are considered as 1.

https://drive.google.com/open?id=138kt00eoteZYwxSH_tUS_5tv6cJGxZwu

Dear @Mariya_Timotey_Mitev

As this is off topic where it was posted, I took the liberty and created a new topic for discussion.

From the images you are showing and what I can see, you have reached the limits of what the pretrained model can do and would have to train your own model with your own annotated images.

Seeing as this is rather more complicated I would suggest you look around in the labs near where you work or with your Scientific IT department for someone to collaborate to do the training.

Of course, if someone else has a better idea or suggestion, I am sure they will answer here and help out :slight_smile:

1 Like