The Random Forest Tree technique

The ability of the various components of a tissue sample from certain image analysis modules to differentiate depends on their capacity to automatically learn the different tissue and cellular architectures present in an image.  These tools are machine learning-based.

The Random Forest Tree technique is one of them. It enables the creation of a probabilistic decision model (random forest) using a learning base of the different categories making up an image.

The algorithm uses different mathematical criteria enabling a description of the color and architecture of the categories to be separated (tissue or cellular). Once the model is created it can be applied to several images when using the following image analysis algorithms :

The protocol :

The stages of creating an RFT project are as follows :

  • Selection of reference images
    Opening of one or several digital slides in CaloPix on which the different categories of tissues or objects to be identified appear. Using windowing, the user can then select the most representative areas of the slide.
  • Identification of the color plane
    The algorithm carries out automatic color deconvolution, which can subsequently be adjusted by the user. The color plane selected must be sufficiently differentiating in order to be able to view the tissue or cellular categories.
  • Definition of categories
    The different categories of objects or tissues present in the image must be defined and named. The user then designates these directly onto the preselected images using a brush.
  • Selection of categorization criteria
    The matrix offered by the RFT tool enables the size of the analysis filter and the relevant categorization criteria to be defined.
  • Categorization verification and adjustment
    After validation, the tissues of the active image are categorized according to the predefined tissue categories. The user checks the categorization mask on several images that have been preselected for the job at hand. The purpose is to maximize the probability of objects belonging to their categories and to reduce uncertainty by adding new markers or modifying parameters.

Once created the RFT project is saved so it can be used in image analysis modules.