Abstract: A Multi-scale Attention Wavelet Network (MAW-Net) is proposed to solve the problem of insufficient robustness for segmentation of lesions of different scales in ultrasound breast images. Through the design of two lightweight network modules, multi-scale splicing module and skip connection up-dimension module, the goal is to integrate rich features and global context information on different scales, reduce the semantic gap between the encoder and decoder, and adapt to the purpose of disease segmentation on different scales. The dual-tree complex wavelet transform is introduced to weaken the influence of noise. The Dice coefficients are 91.32% and 84.23% respectively when tested on two common breast ultrasound data sets UDIAT and BUSI. Compared with other six advanced image segmentation methods, it has the advantages of strong segmentation robustness and small noise impact.