Cell segmentation challenge: DAPI & Transmission

Dear Machine Learning Experts,

@romainGuiet, @oburri, @iarganda, @mweigert, @fjug

Do you know whether there is something useable in Java that would allow me to segment cells in below images? The intensity gradient in the transmission image could be avoided during acquisition, but if the algorithm was tolerant to that, even better :wink:

As a final output I would need nuclei and cell label masks.

Download raw data

I think even the classic 2D U-net publication contained an example of segmenting cells in transmission images, but I don’t know how to access this from Java.

Thank you for your help and time!

Not sure how well that would work, but it’s in Fiji… even the training.

Might work better when you train on artificially noised inputs… (weird as it sounds…)


What about creating your own U-Net, training it in Colab, and saving it to be used from DeepImageJ? You have all the steps in our tutorial from the Neubias textbook chapter we are preparing.


Do you also offer a mode that only uses the three class (bg, boundary, object) U-Net without the denoising?

1 Like

Hi @Christian_Tischer,

I’m blushing to be tagged as a Machine Learning Experts :blush: I was not expecting that :sweat_smile:

My first option would be to try with CellPose, pretty sure the resuls will be amazing
BUT you want something usable in Java so it leaves 3 options that I’m aware of:

  • have you tried ilastik already ? :sweat_smile: it might work for the nuclei part, not sure about the cell body, you may have to homogenize the image first.
  • @oburri and I did some test and hijacked CSBDeep/CARE notebooks to train a model for in silico channel prediction (see below) and then you can use the CSBDeep FIJI plugin.

A Phase to DAPI by @oburri

A HCS_cellMask -> DAPI + Phallo by me

  • I support @iarganda with collab + DeepImageJ, which I also tested. The only issue I encountered was the space limit on googlDrive