Training StarDist on MultiChannel images

Thanks for developing StarDist Uwe Schmidt et al.

Currently in our centre, cell segmentation is done manually by experts and they use multiple channels (usually dystrophin and VDAC) context to identify 1) All the viable cells in the sample & 2) boundaries of these cells

I am trying to build a cell segmentation machine learning model that can segment all the “viable” cells in IMC (image mass cytometry) and EM images of muscle biopsies.

For this, I trained the 2D StarDist model on 29 single channel IMC (image mass cytometry) images of muscle fibres + 470 images from DSB2018. We found this trained model was good at predicting the boundaries of cells but not good with filtering non-viable cells.

I can see the obvious reason for this i.e. the model is not trained on multiple-channel. I understand currently StarDist doesn’t support multi-channel/stacked images as input. But do you have advice on how to deal with this?

if need be I am ready to spend time on modifying StarDist to achieve this objective.

many thanks
PS, I am a trained machine learning engineer

Hi @Atif_Khan

I am not entirely sure what is meant by “viable cells” (maybe you could show an image demonstrating the problem?), but

stardist does support multi-channel input when training a custom model. So maybe training with multichannel input might help?

All the best,