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

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