Apeer segmentation results

Hi Apeer users,
I’m trying to segment an H&E dataset of muscle tissues.
For whatever reason, after training (2 images) and running the segmentation, it appears to only have a mask of the background; where the other results just seem like the normal image.

Am I missing something?

Raw (example 1-of-9):
20210301_MyScan_Slide1_1.tif (9.0 MB)

Result screen on Apeer:

What I see on the web viewer (nothing):

The only result image I can view with some segmentation visible (I assume it BG segmentation):

What is missing for me to get a similar image (segmented on ilastik)?

The goal I to segment single fibers, wherever the tissue is properly sectioned (circular fibers are good, elongated are bad). I don’t care to have bad segmentation as long as I can screen by size and circularity.


  • What stops you from proceeding?
    apeer interface is not so intuitive o me.

  • What have you tried already?
    ilastik segmentation.

Thanks to the helpers!

Hi @Daniel_Waiger

there can be different reasons, why the result is not what you think it should be. Here some general remarks:

  • From the image it looks like you want to train a network for semantic segmentation
  • Did the training work ok? How did it perform? How many classes did you use?
  • a direct comparison with ilastik pixel classification might be tricky, because on APEER you train an EfficientNet, which is different
  • despite the fact that we use many data augmentations two images for training are certainly on the low side

But bottom line is the APEER support team would need to check you training workflow and the actual segmentation workflow as well to be able to really help you. You will need to share your training results and your workflow with them. It should somehow a bit like this …

Therefore I recommend to directly use the “Contact” option on APEER or ask your question on the APEER Forum.

Remark: When APEER was launched we actually asked this community if APEER could become a partner on this forum as well because the module code can be public), but after a somewhat heated discussion we were not allowed to join, because APEER is obviously owned by ZEISS …
This is when we had to start our own forum.

On a personal note i think it would be much better for everybody if this decision would have gone the other way …

1 Like

Actually, I think I went for instance segmentation. Might it be the reason?

I had 4 classes. For the initial segmentation I don’t care about accuracy (2 images are indeed a small training dataset), I just wanted to try the tool on a relatively hard dataset.

Hi @Daniel_Waiger,

on APEER can train both variants, DNNs for instance segmentation or sematic segmentation. And later on you have to select the correct class for the segmentation. Maybe you selected the wrong one for the segmentation?

But check first what you selected here (screener from one of my trainings) → I wend for semantic segmentation and labelled one class (“nuc”) only. This means all other pixels in your training images will be consiered background

As I said, please contact the APEER support team directly for this.

1 Like

Thanks, I’ll look into that.