Deep learning training model

Hi there,

I need to track the following cells migration, but I still didn’t manage to segment them.
I’ve been trying to use Stardist (@uschmidt83 ) and cellpose but I need a model that better adjust to these images.

testing_t24_0080.tif (2.0 MB) testing_t24_0077.tif (2.0 MB)

Do you know if it already exists? If not, do you have suggestions on how to do it from scratch?
Where and how can I label the cells (the ground-truth images)? What is the recommended image size and number of images to create the training model?

Thank you in advance


1 Like

I get these with trainable weka segmentation. Are they better than your current segmentations?

…77 segments using the model of model developed with …80

…80 segments using the model of model developed with …80

Segmenting touching objects is hard…good luck

Hi @mmvpgs

You could potentially do it with classical image processing, but that might be trick since (I assume) you want to individually segment all the cells (which is challenging when they’re touching, as @rondespain mentioned).

I think StarDist would work for these images. Please take a look at this about annotating images. (Alternatively, you could use ImageJ and the Roi Manager to annotate each cell.)

Do all your images look like this?

The two you posted seem to come from a timelapse from nearby time points (they are very similar). I’d suggest to annotate complete frames from spread-out time points, i.e. don’t choose images to annotate that look too similar. You could start labeling 10-20 frames, train a model, and check how the results look. Of course, more labeled training images are always better.


1 Like

I tried multi template matching and it looks like it might be helpful to you.

training videos

1 Like