Abstract: Accurate segmentation of breast masses is of great importance in the process of computer aided diagnosis for breast cancer screening. However, the image gray of breast mass in X-ray imaging is close to the background and its shape is also irregular, which makes a great challenge to achieve accurate segmentation. In order to further improve the segmentation performance, a novel breast mass segmentation method based on spatial adaptive and mixed loss adversarial network is proposed. Firstly, a separable convolutional U-Net model is proposed as a generator in adversarial network to reduce the computational cost of parameters and time complexity. Secondly, the spatial adaptive normalization layer is added to the discriminator to obtain the semantic information contained in the segmentation mask. Finally, considering the influence of category imbalance, semantic consistency and other factors, a mixed loss function is proposed to integrate the admissive loss, segmentation loss and perception loss, so as to improve model learning effect. Rigorous experimental on INbreast and CBIS-DDSM breast image datasets show that the proposed method has achieved accuracy of 99.35% and 99.72%, dice of 81.27% and 82.01%, respectively. The experimental results demonstrate that the segmentation performance of the proposed model outperforms the existing methods.