A Novel Multi-Scale Adversarial Networks for Precise Segmentation of X-Ray Breast Mass

With the constant changes of people's lifestyle and living environment, the morbidity of breast cancer is increasing year by year. It is highly imperative to develop an effective breast mass segmentation method for early breast cancer diagnosis. However, segmenting breast masses in mammograms is still one hot issue with enormous challenges because of masses' irregular shapes and various sizes. In this study, we propose multi-adversarial learning to capture multi-scale image information for accurate breast mass segmentation. To effectively reinforce higher-order consistency in the segmentation results, the proposed network introduces the idea of adversarial networks, mainly consisting of a segmentation network and a discrimination network. An improved U-Net is introduced as the segmentation network to generate masks of the suspicious regions, while the discrimination network combines three convolutional critic networks that operate at different scales to discriminate the input masks. To weaken the unbalanced class problem and produce fine-grained segmentation results, weighted cross entropy loss and Earth-Mover distance are jointly used as an integrated loss function to guide the optimization process. Furthermore, the spectral normalization is adopted to the critics to alleviate the instability of training. The effectiveness of the proposed method is evaluated on two public datasets (INbreast and CBIS-DDSM). Experimental results empirically demonstrate that our method outperforms FCNs-based methods and the state-of-the-art method with dice of 81.64% for INbreast and 82.16% for CBIS-DDSM.