What is minimum cross entropy thresholding good for


I read the description available in cell profiler 3.0 help but I was hoping someone could further explain the use of minimum cross entropy thresholding, i.e., in what situations is it likely to outperform Otsu and why? (Or outperform MCT thresholding for that matter).



It was the product of a few overlapping conversations:

  • We felt like we added more thresholding methods than we could reasonably maintain.
  • We wanted to use the thresholding methods available from scikit-image to simplify maintenance and provide some structure around deciding whether we should include new thresholding methods (i.e. we’d consider adding a method if it’s already been contributed to scikit-image).
  • We wanted to use thresholding methods that were familiar to users like manual thresholding or Otsu’s method.

We chose Li’s method as a default for two basic reasons: performance is excellent and it’s familiar to ImageJ users since it’s one of the most popular algorithms from ImageJ’s excellent Auto Threshold plug-in for thresholding cellular objects (it’s also extremely easy for users to find more information about the method because it’s heavily used).