A Dual Feature Selection Approach for Classification of Breast Tumors in Ultrasound Images Using ANN and SVM

Appropriate selection of features to represent breast patterns is of important concern in breast tissue classification. The traditional approaches of feature selection using one evaluation criterion have shown limited performance in decision support due to their biasness towards single criterion. In this paper a novel feature selection approach based on dual evaluation criteria is proposed. Database of 89 breast ultrasound images were used in the experiments. The acquired images were first subjected to wavelet based despeckle filter to reduce speckle noise. Then, total of 457 texture and shape features were extracted from despeckled images. Out of 457 features, 12 most relevant features were selected using proposed feature selection approach. To evaluate the selected features, ANN and SVM classification models were used for classifying benign and malignant breast tumors. Five performance measures namely accuracy, sensitivity, specificity, AUC and MCC are used to compare the performance of classifiers and feature selection techniques. Results show that selection of relevant features using more than one evaluation criteria improves the performance of classifier as compared to that using only one criterion with reference to most of the performance measures.