Breast cancer is regarded as the leading cause of death in females in Vietnam. The screening mammogram is one of the most common methods used for early detection of breast cancer. However, the radiopaque artifacts such as wedges and labels or pectoral muscle in the mammogram images can reduce the rate of accuracy. Therefore, the pre-processing of mammogram images creates the possibility of detection of lesion successfully. In this project, we propose automatic techniques (Deep learning) to preprocess mammograms images and present a new approach to detect and segment mass cancer on screening mammograms for early breast cancer detection stage.