Feature selection and classifier performance in computer-aided diagnosis for breast ultrasound

We propose a feature selection method for classifying breast ultrasound (BUS) images based on mutual information technique and statistical tests. The BUS dataset consisted of 641 BUS images (228 carcinomas and 413 benign lesions) and every image was segmented by a technique based on Watershed transform. Thereafter, 22 morphological features were computed from segmented lesions and the resultant feature space was ranked by mutual information approach, where the first feature presents the largest discrimination power between benign and malignant classes. Next, feature subsets were built by adding iteratively the first m ranked attributes until all of the 22 features were considered. The .632+ bootstrap method estimated the discrimination performance of each feature subset in terms of the area under the ROC curve (AUC), by using the Fisher discriminant analysis (FLDA) as classifier. The results pointed out that the AUC values were 0.952 and 0.953 for the reduced (with seven features) and complete sets (with 22 features), respectively. Hence, dimensionality reduction was reached while maintaining the classification performance.