Breast tumor segmentation with prior knowledge learning

Abstract

Automatic breast tumor segmentation is a crucial step for breast ultrasound images analysis. Prior knowledge can be used to improve segmentation performance. However, commonly used prior information such as intensity, texture and shape may be useless due to complicated characteristics of breast tumor in ultrasound images. In this paper, we propose a novel prior knowledge of the abnormal tumor regions, which may be complementary to base segmentation model. Based on this idea, we develop a breast tumor segmentation approach with prior knowledge learning. The proposed method mainly consists of two steps: prior knowledge learning and segmentation model construction. In the first step, prior knowledge learning model is developed to learn prior information which can be used to classify abnormal tumor regions correctly. It's difficult for base segmentation model to obtain accurate segmentation result of abnormal tumor areas. Therefore, learned prior knowledge is complementary to base segmentation model. In order to exploit learned prior knowledge, prior knowledge-based constraints are incorporated into the base segmentation model for robust segmentation model construction. In order to verify performance of the proposed method, we construct a breast ultrasound images database contained 186 cases (135 benign cases and 51 malignant cases) by collecting the breast images from four types of ultrasonic devices. Our experimental results on the constructed database demonstrate the effectiveness and robustness of the proposed method.