Cell Cluster Segmentation

Segmenting cells in cell clusters is very challenging because they cannot be separated easily from the background – thus classical segmentation algorithms fail.
Unfortunately in WSI (whole slide imaging) you will have to deal with these situations – especially if you don’t have a single cell layer, like e.g. in tumor tissue.

A collaboration with the  ZHAW  (Zürcher Hochschule für Angewandte Wissenschaften) lead to a new Mumford-Shah based segmentation algorithm which solves this problem and is able to deal with cell clusters.

The new Orbit version 2.65 integrates this solution as a new option in the segmentation options dialog (F3 -> tab segmentation). There you can activate the Mumford-Shah segmentation algorithm:

which will then be used instead of the default segmentation algorithm. The basic segmentation process (training a foreground/background model and setting this as a primary segmentation model) remains.
Two options will influence the cell detection and splitting:

  • Object size: Default cell size in pixels. This parameter defines the object splitting based in object size. Smaller values will split objects more frequently.
  • Intensity split: Smoothening factor. This parameter defines the object splitting based on object intensities.  Smaller values will split objects more frequently.

Here is an example:


segmented image

original image

 

I want to thank Remo Schweizer, Benjamin Meier and their supervisors for this great contribution. It is a fantastic example how academia and open source projects can profit from each other!