Label-Distribution Learning-Embedded Active Contour Model for Breast Tumor Segmentation

Abstract

Tumor segmentation is the foundation of breast ultrasound image analysis. However, intensity inhomogeneity occurred in ultrasound images results in the ambiguous segmentation. In order to tackle the challenge, this paper proposed label-distribution learning embedded active contour model for the breast tumor segmentation. Considering that reasonable exploitation of label ambiguity may help to improve performance, a deep pixel-wise label distribution learning model is first proposed to learn an ambiguous label map. The learned map is independent on the intensity variation, which is robust to the intensity inhomogeneity. After that, a novel label distribution learning embedded active contour model is proposed. The new energy function is developed by introducing the new label distribution fitting energy into the active contour model framework. The proposed new fitting energy can enforce the label of pixels to be similar to the learned distribution map, which improves the robustness to the intensity inhomogeneity. To demonstrate the effectiveness of the proposed method, we conduct the experiment on the breast ultrasound images database which consists of 135 benign cases and 51 malignant cases. Our experimental results demonstrated that the proposed method outperforms the state of the art.