Spatial Enhanced Rotation Aware Network for Breast Mass Segmentation in Digital Mammogram

Breast cancer is the most common with highest mortality risk among female worldwide and breast mass effective sign for identification. Thus, accurate segmentation of regarded as a key step to reduce death rate. Traditional methods require prior knowledge manually set parameters, while recent studies prefer construct neural networks based on feature reuse. However, can display in different orientations spatial context complex, which makes remain challenging task. For these concerns, we propose Spatial Enhanced Rotation Aware Network (SERAN) automatic segmentation. SERAN consists two critical components: 1) residual attention encoder enhancement mechanism extraction, 2) decoder constructed by multi-stream rotation aware blocks fusion prediction refinement. To optimize better avoid misclassification background area, regulation item named Inside-outside Loss (IOL) used training procedure. The experimental results tested representative subset Digital Database Screening Mammography (DDSM) dataset show that outperforms state-of-the-art adopted evaluation metrics.