Ensemble features selection method as tool for breast cancer classification

Abstract

This work aims to gather experimental evidence of features relevance, as well as finding a breast cancer classification scheme that provides the high performance over the area under receiver operating characteristic curve (AUC). An ensemble feature selection method (named RMean) based on the mean criteria for indexing relevant features is presented. The proposed method provided better classification performances (statistically significant) than those who constitute the baseline, attaining AUC scores of 0.7775 with the support vector machine on microcalcifications dataset and 0.9440 with the feed-forward-backpropagation neural network classifier on masses dataset. The most relevant features for microcalcifications classification were: mammographic stroma distortion, density, right bottom quadrant, perimeter, standard deviation, entropy, and angular second moment. Meanwhile, to classify masses were: mammographic stroma distortion, mammographic calcification, mammographic nodules, density, circularity, roughness, and shape.