A feature selection methodology for breast ultrasound classification

In this paper we proposed a feature selection methodology for classifying breast ultrasound (BUS) images based on principal component analysis (PCA) and mutual information (MI). The BUS dataset consisted of 641 BUS images (228 carcinomas and 413 benign lesions). Besides, three M-dimensional feature sets were built: morphological (M = 22), texture (M = 502), and combined (M = 524). These sets were ranked by PCA and MI approaches, where the first feature presents the largest discrimination power between benign and malignant classes. Next, m-dimensional feature subsets (where m <; M) were created by adding iteratively the first m attributes. The .632+ bootstrap error was estimated at each iteration by using the Fisher discriminant analysis (FLDA) as classifier. Thus, at the argument of the minimum of the error curve is found the best m-dimensional feature subset. Finally, the area under ROC curve (AUC) was used as figure of merit to evaluate the discrimination power of selected features. The results pointed out that the best classification performance was reached by the “combined-MI” set with AUC = 0.951 and 13 features, whereas the “combined-complete” set attached AUC = 0.657 with 524 features.