Automatic segmentation of breast masses in DBT images based on RMAU-Net

Accurate breast mass segmentation is important for the diagnosis and treatment of early breast cancer. Digital breast tomosynthesis (DBT) has been widely used for breast cancer screening with a high detection rate for lesions. However,the high breast densities and low contrast in DBT images make the automatic segmentation of breast masses very challenging. In order to efficiently and accurately segment the masses in DBT images,this paper proposes a residual multi-attention U-shaped segmentation network (RMAU-Net),which utilizes a residual structure to avoid performance degradation caused by gradient vanishing. Meanwhile,a deep attention feature fusion module and a multipath high-level feature fusion module are used in the network to improve the feature extraction ability of the network as well as the ability to recognize the boundary of suspicious regions.The RMAU-Net performs segmentationon a private DBT image dataset (DBT_SZ) and achieves a Dice of 86.77%,a sensitivity of 87.84%,and an IOU of 80.15%. In addition,this paper compares RMAU-Net with some advanced segmentation networks.Experimental results show that RMAU-Net can extract mass edges more accurately so that improve the segmentation accuracy.