Automatic detection of ultrasound breast lesions: a novel saliency detection model based on multiple priors

Due to the complex tissue structure of the breast, breast ultrasound (BUS) images exhibit the characteristics of low-contrast, lesion boundary blurring. Therefore, accurately automatic detection of ultrasound breast lesions is an important challenge for a computer-aided diagnosis system. Although previous saliency detection methods have been applied to detect breast lesions, they do not take full advantages of the prior knowledge of breast lesions. Here, to further accurately detect the breast lesions, a novel saliency detection method is proposed for BUS images, which seamlessly incorporates multiple priors into a hybrid architecture. To reduce the speckle noise and enhance the contrast, the BUS images are preprocessed by the methods of median filtering and a proposed adaptive thresholding. Also, to reveal the differences of benign and malignant lesions, a heat map based on the boundary of the breast lesions is established. Extensive experiments indicate that the proposed saliency detection method achieves an excellent performance of 0.925 accuracy, 0.871 sensitivity, 0.889 dice, and 0.912 F-measure on breast lesions detection in the BUS images, which is superior to the saliency detection models with a single prior. The boundary heat maps of the lesions also visually reflect the differences between benign and malignant lesions, which may potentially be used for automated computer diagnosis to assist radiologists in detection and identification of breast lesions.