Multi-stream Information-Based Neural Network for Mammogram Mass Segmentation

Mass segmentation is the first step in computer-aided detection (CAD) systems for classification of breast masses as malignant or benign, and it greatly impacts the accuracy of CAD systems. This paper proposes a model called region-based graph convolution and the atrous spatial pyramid pooling network (RGC-ASPP-Net), by considering mass context information, such as the features of location and size of mammogram masses, to yield better segmentation results for the CAD systems of mammogram diagnosis. Specifically, it introduces ASPP module in its skip-connection layer, to capture multi-scale mass context information. Then, it constructs a graph convolution module based on the clustering results of mass positions, for taking factors of the location of mammogram masses into account during the process of segmentation. We evaluated our model on the CBIS-DDSM dataset for conducting segmentation tasks, and the results demonstrate that our model RGC-ASPP-Net outperforms PSPNet, DeepLabV3+, AUnet and ASPP-FC-DenseNet by a large margin in terms of segmentation performance.