Article Preview
TopIntroduction
Region-based level sets have been considered as an elective method for breast ultrasound image segmentation. Their main advantages are insensitive initialization (independent of the position of the initial contour) and fast computation. This method is an active contour evolving in time due to the measurement of intrinsic geometric in images. The evolving contours commonly merge and split, allowing several objects to be simultaneously detected in both exterior and interior boundaries (Vicent et al., 1993). Various region-based level sets were used to extract features, for example, localizing region-based active contours-LRBAC (Lankton & Tannenbaum, 2008), fuzzy clustering, and lattice Boltzmann method-FCLBM (Balla-Arab, Gao, & Wang, 2013), and online region-based active contour model-ORACM (Tulu, 2013). The mentioned methods were applied in this work for features extraction in a breast ultrasound image.
The fusion method called FM (Keatmanee et al., 2019) was contributed to extracting significant information from three ultrasound imaging modalities, which are US (conventional ultrasound), Doppler, and elasticity images. The essential features which were related to the fusion of the three images are dark gray regions in the conventional ultrasound image, hard tissue parts in elasticity, and vascular flow areas in Doppler. According to the benefits of the supplement ultrasound imaging modalities, the union of them could improve the accuracy of radiologists’ diagnosis in terms of specificity and sensitivity (Cho et al., 2012). Hence, it could increase the accuracy of breast cancer contour detection using active contour models as well. FM aims to locate the common area of the essential features extracted from the three images using distance transform. First, it could be done by the thresholding method to generate a binary image. After that, the boundaries of all dark regions in the binary image could be traced. The prospective boundary of breast cancer contours could regard these characters; it should share a common area with the Doppler and elastic region; its center should be close to the center of the Doppler region.
The segmented contours obtained from the region-based level sets could be considered as over-segmentation for breast cancer contour detection. Usually, there is only one abnormal hypoechoic lesion appearing in the US image. Consequently, the output contours from level sets may not be satisfied. Therefore, we introduce a binary classifier using Support Vector Machine (SVM) to distinguish a correct contour from false segmented contours. Features of the proposed model were obtained not only from the level sets but also from the FM. Consequently, the level sets extracted prospective contours from the US, and the FM was utilized to find meaningful features such as the center and the common area obtained from the fusion of the three ultrasonography images. The model aims to predict whether the segmented contours by level sets were breast cancer contours or false contours.
This paper proposes a breast cancer contour detection model applied in conventional US, elasticity, and Doppler images. The SVM was used as a binary classification model. The features for training, validating, and testing were extracted from two methods which were region-based level sets (Lankton & Tannenbaum, 2008; Balla-Arab, Gao, & Wang, 2013; Tulu, 2013) and the FM. The proposed model was developed to detect a single contour which was a boundary of the breast tumor in the conventional US. Depicting in Figure 1, the input image (conventional ultrasound) was preprocessed for noise suppression and edge detection. After that, the edge maps were captured by the level sets method. The detected edges combined with features from FM were passed through the binary SVM for breast cancer contour classification. Whereas, the process of features extraction performed by the level sets and the FM was done separately.