Breast Density Segmentation in Mammograms Based on Dual Attention Mechanism

In response to the problem that poor segmentation accuracy results from artifacts in the mammogram, this paper proposes combining the U-Net with Coordinated Attention and Attention Gates to enhance target feature information and suppress irrelevant regions. First of all, the INbreast dataset is preprocessed to remove external artifacts. Second, in the contracting path, enhance features of the Region of Interest(ROI) of the mammogram through the Coordinate Attention Module. Finally, in the expansive path, the local feature enhancement can be achieved by Attention Gates is used instead to combine the shallow layer and upsampling feature maps directly. The experimental results show that the proposed algorithm has a good segmentation effect for the mammogram, and its the Dice Similarity Coefficient (DSC) and the Intersection of Union(IoU) are 91.8% and 85.8%, respectively. Furthermore, we obtained DSC and IoU of 98.4%, 96.8%, respectively, for women with high breast density. Compared with the conditional Generative Adversarial Networks (cGAN) algorithm, the DSC increased by 3.36%, IoU increased by 5.91%. The better segmentation achieved can help doctors accurately judge breast density categories.