Fusion of SFS-SVM feature selection methods using robust rank aggregation for optimal feature subset selection for mammogram classification

Feature selection and classification plays an important role in the design and development of a computer aided detection and diagnostics (CAD) tool for breast cancer detection from mammograms. In literature, the various feature selection methods exists such as filter based, wrapper based, and hybrid methods whose aim is to select the most relevant and minimum redundant features from the extracted feature set. The various feature selection methods are based on various basis criteria on which they are designed;and they produce different feature sets even for the same dataset with varying performance measures for a chosen classifier. Therefore, the basic question arises that which one is better and if all of them performs better with slight variations in their performance measures then what should be the robust method to select the optimal feature subset associated with constant and better performance for a chosen dataset? Hence, in this paper, the above mentioned issues are addressed by proposing a wrapper based feature selection method based on sequential forward selection (SFS) and support vector machine (SVM) for optimal feature subset selection. The proposed method is derived from the fusion of SFS-SVM feature selection methods for various kernel spaces of SVM i.e. the feature sets obtained from SFS-SVM-Linear, SFS-SVM-RBF, SFS-SVM-Quadratic, SFS-SVMPolynomial, and SFS-SVM-MLP wrapper based feature selection methods are fused using robust rank aggregation method. Finally, an optimal feature subset is obtained from the fused feature set by evaluating the minimum misclassification error of a chosen classifier.