A Comparison of Feature Selection Methods for the Detection of Breast Cancers in Mammograms: Adaptive Sequential Floating Search vs. Genetic Algorithm

This paper presents a comparison of feature selection methods for a unified detection of breast cancers in mammograms. A set of features, including curvilinear features, texture features, Gabor features, and multi-resolution features, were extracted from a region of 512×512 pixels containing normal tissue or breast cancer. Adaptive floating search and genetic algorithm were used for the feature selection, and a linear discriminant analysis (LDA) was used for the classification of cancer regions from normal regions. The performance is evaluated using Az, the area under ROC curve. On a dataset consisting 296 normal regions and 164 cancer regions (53 masses, 56 spiculated lesions, and 55 calcifications), adaptive floating search achieved Az=0.96 with comparison to Az=0.93 of CHC genetic algorithm and Az=0.90 of simple genetic algorithm.