ML-based Filipodia detector

Dear All,

I am looking for a reliable, robust 2D “Filipodia Detector”.

I have not tried, but i guess a simple Random Forest approach would not be reliable enough for our data.
(see attached example)

Autocontext might be an option, but a NN-based approach would likely work best.

(ideally with a pre-trained classifier).

(Zoom in 1 shows the filipodia, Zoom in 2 the label in white)

Do you have any recommendations ?

Thanks a lot & Kind regards


@constantinpape @mweigert @Christian_Tischer

Hey @Tobias,

have you given ilastik (pixel classification, autocontext, potentially with object classification afterwards for cleanups) a try? But of course you are right that getting the multiple lines you have highlighted in your zoom-in might be challenging.

Dear @k-dominik,

no i have not. Do you think we should??

I have way to many project at the same time to do it myself, and I had too strong doubts (see below) to encourage our user to do so.

We have no “free” ground truth, as they quantify so much by hand that they just count them (not draw).

Moreover the same classifier shall work on control and various mutants, where filipodia might vary.
And it is from a tissue whole mount, thus the staining might also vary from batch to batch.

Thus, i had hoped that something is out there for filipodia we could just use.
(or re-train, but being pretty sure it would work in the end…).

Kind regards


Hi @Tobias,

I don’t know of any of the shelf pre-trained network for this task. So I would definitely try with ilastik first as well; even if the results are not satisfactory yet, this can be used to bootstrap the necessary training data.

Re ilastik: the data looks a bit too difficult for normal pixel classification, but it might work well with Autocontext.

Also, what exactly do you want to measure? Depending on this different methods might be useful. Do you just need the total area of filipoidia, the number of individual filipodia or even properties of individual filipodia like length, area etc?

Dear @constantinpape,

Thanks a lot!

Until now they just count filipodia by hand.
But more measures might be useful.
The general interest of the lab is the characterization of alterations in vascular morphology.
For all other measures they are interested in i found tools i could just adapt, combine & make turn-key.

For the filipodia, i told them, “sorry this is not possible with the current state of the art”
(but wanted) to make sure, i did not miss any “Fili”-pose or “Fili”-Dist pre-print :slight_smile:

I will tell them, that they could try the approach suggest by you.
(and we can maybe talk in person about details on Weds??).

Hopefully soon someone from the “segment dense and denser nuclei” community
considers paying more attention to the “leading edge” :slight_smile:

Kind regards


Just to clarify, I definitely think this task is feasible, but as with most things, the question is if it’s worth to invest the time to set up the automatic approach.
Happy to discuss this in person on Wednesday.