Improving classification performance of breast lesions on ultrasonography.

Several morphological and texture features aiming to distinguish between benign and malignant lesions on breast ultrasound (BUS) have been proposed in the literature. Various authors also claim that their particular feature sets are capable of reaching adequate classification rate. However, there are still several features that have not been tested together for determining the feature set that effectively improves classification performance. Hence, in this paper, we compiled distinct morphological and texture features widely used in computer-aided diagnosis systems for BUS images. A total of 26 morphological and 1465 texture features were computed from 641 BUS images (413 benign and 228 malignant lesions). A feature selection methodology, based on mutual information and statistical tests, was used to evaluate the discrimination power of distinct feature subsets. The .632+ bootstrap method was used to estimate the classification performance of each feature subset, by using the local Fisher discriminant analysis (LFDA), with linear kernel, as classifier, and the area under ROC curve (AUC) as performance index. The experimental results indicated that the best classification performance is AUC=0.942, obtained by a morphological set with five features. In addition, this morphological set outperformed the best texture set with four features, which attained AUC=0.897. The classification performances of 11 feature sets proposed in the literature were also surpassed by such morphological feature set. Highlights1491 features are evaluated for classifying breast lesions on ultrasound.Feature selection is based on mutual information and statistical tests.5 morphological and 4 texture features achieve the best classification performance.11 feature sets from the literature are surpassed by the 5 morphological features.