Allow the algorithms to perform optimization over the model parameters (kernel, C and gamma for SVM).

See: Project-Platypus/Platypus#31

Same as #9, but using hypervolume as performance metric.

See: https://platypus.readthedocs.io/en/latest/experimenter.html

Implement a baseline SVM classifier for radiomics and deep_radiomics sets.

Basic steps:

• Read data
• Preprocessing: remove columns, normalize features
• Validation: 10-fold cross-validation
  • Training
  • Evaluation (accuracy, sensitivity, specificity, F1-score, ROC and AUC)

Plot a graphic of fitness over generations.

Possible obstacles:

Platypus uses evaluations instead of generations

algorithm = NSGAII(SVM(), population_size=10)
algorithm.run(100)

Can we use evaluation/populations_size as the generation number? Investigate source code

Can we collect the fitness during execution?

I don't know if it is possible to obtain the fitness over the generations or if we need to do something like:

algorithm.run(10)
algorithm.run(20)
...
algorithm.run(90)
algorithm.run(100)

Check https://github.com/manuel-calzolari/sklearn-genetic

Current Platypus implementation allows dominated solutions to feature on the archive:

The simplest solution would consist of a filter function that removes the dominated solution.

The analysis of features consists of evaluating which radiomics features are more common in models with higher sensitivity or specificity.

The steps are:

1. Given a threshold, select only models with values higher than the threshold
2. Count the occurrence of each feature in each model
3. Compute the percentage of occurence (count/number of models)
4. Plot a feature importance chart with the feature names and values of importance.

Example: