I need suggestions for automatic segmentation

Hi guys,

I need suggestions to solve a segmentation problem.
The propose is to count fibres that cross the intradermal layer as suggested in figure result. I was first trying to segment the intradermal using either ch2 (dapi) or ch3 background and then find the maximum values to identify the fibres in channel 1.
The problem is with the segmentation od intradermal layer. Do you have any suggestion that could be generalized for other similar images?
I also have the option of stitching all the images of the same biopsy in order to better identify the intradermal, but it turns to be much heavy for processing

Thank you in advance.

Mafaldaresult.tif (3.0 MB)


Would you be able to help us help you better by defining better what you need/want to segment? We all come from different backgrounds… so not sure where the ‘intradermal layer’ is etc. Just avoid the jargon and show us if you can (just a manual delineation as an example image) what you aim to segment - who is in what channel, etc.

Too - here are some helpful links on Segmenting in ImageJ (and also Scripting in case you want to batch process):

1 Like

Hi, I’ve attached a result.tiff file in the previous post with the manual delineation of what is the intradermal site (Roi1). In the three channels is a site of transition from lighter to darker region. The idea is to count the fibres (Roi 2) that crosses that ROI1.

Thank you for the links. I know how to segment, I just need ideas for good and automatic segmentation in this case. Like, if you suggest segmentation by thresholding or by deep learning, etc.



My bad! Missed that annotated image… but your extra description was really helpful in any case. Here’s a snapshot of it to help others see what you wish to segment:

So… a few issues. Looks like your stitching is a bit ‘off’:

Will you process each tile individually or do you want to do the whole stitched image? This will obviously change the segmentation workflow you decide to build…

How are you currently carrying out your stitching??

You can give Trainable Weka Segmentation a try… it’s linked above. Using the following settings:

I got these results:

It’s not too bad - though room for improvement. It’s true that with ‘classic’/ad hoc methods for thresholding/segmentation - you’ll have issues. So machine learning might be a good place to start.