Ilastik, eccentricity, object classification

Hello there,
I was wondering if I could possibly have some feedback or help. I am working with SEM images of nanoparticles, in attempt to separate the particles by them being a monomer, dimer, trimer, or larger polymer.

Below is an SEM image of some nanoparticles on an optical fiber

F3_35K_002.tif (6.4 MB)

From using the Pixel and object classification in ilastik I attempted to separate the particles by them being a monomer, dimer, trimer, or polymer as seen below, using the “principal component objects, radius, convenxity, # of defects, branch length, and # of branches” features in ilastik and produced the image below.

export.tiff (14.3 MB)

Below is a prediction from a different image from ilastik

export2.tiff (18.0 MB)

This one worked out decently, but at different magnifications in the SEM or sometimes different images, the prediction starts to be not as good. As seen below

export3.tiff (9.4 MB)

I have been working to resolve this issue, but to not much success. Is there a known way to classify objects based on their eccentricity, or maybe a way to classify its shape that has less to do with size?

Hello @Daniel_Piatigorski,

first of all, welcome to the community :slight_smile:

For starters I just want to establish, in case you are using the combined workflow, that this is not recommended. You should always split the two steps in one pixel classification and one object classification project. We keep this workflow around to give quick demos, but in production there are some issues with it.

Could you, by any chance, correct for the different magnification level by rescaling the data to a common resolution? Size sounds like a pretty decisive factor in the classification problem.

If rescaling is impossible for some reason, from the features you have listed, you could try to exclude all that imply that some length is used, like the radii, branch length… Here you can find a video where we explain the available object features a bit, that could simplify choice of features a bit.