Multi-Scale Fusion U-Net for the Segmentation of Breast Lesions

Abstract

Breast lesion is a malignant tumor that occurs in the epithelial tissue of the breast. The early detection of breast lesions can make patients for treatment and improve survival rate. Thus, the accurate and automatic segmentation of breast lesions from ultrasound images is a fundamental task. However, the effectively segmentation of breast lesions is still faced with two challenges. One is the characteristics of breast lesions’ multi-scale and the other one is blurred edges making segmentation difficult. To solve these problems, we propose a deep learning architecture, named Multi-scale Fusion U-Net (MF U-Net), which extracts the texture features and edge features of the image. It includes two novel modules and a new focal loss: 1) the Fusion Module (WFM) which segmenting irregular and fuzzy breast lesions, 2) the Multi-Scale Dilated Convolutions Module (MDCM) which overcoming the segmentation difficulties caused by large-scale changes in breast lesions, and 3) focal-DSC loss is proposed to solve the class imbalance problems in breast lesions segmentation. Moreover, there are some convolutional layers with different receptive fields in MDCM, which improves the network’s ability to extract multi-scale features. Comparative experiments reveal that the MF U-Net proposed in this paper outperforms other segmentation methods, and the proposed MF U-Net achieves state-of-the-art breast lesions segmentation results with 0.9421 Recall, 0.9345 Precision, 0.0694 FPs/image, 0.9535 DSC and 0.9112 IOU on Benchmark for Breast Ultrasound Image Segmentation (BUSIS) dataset.