Evaluation of Features Selection Methods for Breast Cancer Classification

In this work are evaluated features selection methods for breast cancer classification in segmented mammographic lesions using two categories of extracted features (numeric and nominal). Numeric included statistical, shape and texture lesion descriptors, and nominal are related with patient’s associated metadata descriptors (clinical history). Datasets of features vectors were created using six different approaches: Correlate-based Feature Selector (CBF) in association with three heuristic search strategies (Best First (BF), Greedy Stepwise (GS) and Genetic Search Algorithm (GSA)), Chi Square Discretization, One Rule and RELIEF methods. Selected datasets were classified using Feed Forward Backpropagation Neural Network (FFBP), Support Vector Machine (SVM) and Decision Tree J48 (DTJ48) based Machine Learning classifiers (MLC) models for a comparative performance evaluation. The tests performed allowed to identify better approaches/combinations of features subsets and classifiers for breast cancer classification methods.