Table I.

Processing scheme alternating supervised and non-supervised learning

1. Establish training sets

2. Feature selection

3. Supervised learning

4. Classification rule

a. Plot of score distributions

5. Processing to check for heterogeneity

a. choice of features for non-supervised learning (educated      guess)

b. non-supervised learning algorithm

c. statistical significance test for subpopulations

6. Processing of subpopulations

a. plot of confidence ellipses for nuclei

b. plot of tolerance ellipses for case means

7. Alternative choice :

a. submit subpopulations to formal feature selection e.g.      Kruskal Wallis test or Genchi & Mori ambiguity measure      [18, 32]

8. Non-supervised learning algorithm with targeted features

9. Statistical significance testing of subpopulations

a. Plot of confidence ellipses for nuclei

b. Plot of tolerance ellipses for case means

10. Supervised learning (e.g. discriminant analysis of subpopulations)

11. Classification rules

a. Plot of score distributions