Improving the performance of machine learning classifiers for Breast Cancer diagnosis based on feature selection

This paper proposed a comprehensive algorithm for building machine learning classifiers for Breast Cancer diagnosis based on the suitable combination of feature selection methods that provide high performance over the Area Under receiver operating characteristic Curve (AUC). The new developed method allows both for exploring and ranking search spaces of image-based features, and selecting subsets of optimal features for feeding Machine Learning Classifiers (MLCs). The method was evaluated using six mammography-based datasets (containing calcifications and masses lesions) with different configurations extracted from two public Breast Cancer databases. According to the Wilcoxon Statistical Test, the proposed method demonstrated to provide competitive Breast Cancer classification schemes reducing the number of employed features for each experimental dataset.