Evaluation of Image Features Within and Surrounding Lesion Region for Risk Stratification in Breast Ultrasound Images

Feature extraction and classification plays a crucial role in the automated analysis of breast ultrasound (BUS) images. Due to varying sonographic characteristics of benign and malignant lesions, the texture and shape features are mostly employed for designing computer-aided diagnosis (CAD) systems of BUS images. The existing CAD systems use features that are extracted either from the lesion segmented area obtained through segmentation techniques or a rectangular region of interest (ROI) extracted under the guidance of expert Radiologists. However, the significance of features extracted from region comprising only the lesion area is still little explored. This paper investigates the significance of features extracted from the lesion area, lesion surrounding area and rectangular ROI for classification of BUS images. The experiments were conducted on the database of 294 BUS images (104 benign and 190 malignant). Initially, the acquired BUS images were preprocessed through speckle reducing anisotropic diffusion (SRAD) for speckle noise removal. The preprocessed images are segmented using a hybrid segmentation approach including a combination of region-based active contour driven by region-scalable fitting (RBACM-RSF) model and multi-scale Gaussian kernel fuzzy c-means clustering with spatial bias correction (MsGKFCM_S) for getting ROI confined area. The segmented images were further partitioned into two parts (lesion area and lesion surrounding area). Subsequently, a total of 457 texture and shape attributes were extracted from within the lesion area, lesion surrounding area and rectangular ROI comprising of both lesion and its surrounding area. The significance of these features is evaluated using different classifiers (i.e. support vector machine (SVM), Back-propagation artificial neural network (BPANN), Random Forest, AdaBoost). The results indicate that features extracted from within lesion area achieve a maximum classification accuracy of 98.980% with the lowest computational time when linear kernel-based SVM is used.