Breast cancer is one of the leading fatal disease worldwide with high risk control if early discovered. Conventional method for breast screening is x-ray mammography, which is known to be challenging for early detection of cancer lesions. The dense breast structure produced due to the compression process during imaging lead to difficulties to recognize small size abnormalities. Also, inter- and intra-variations of breast tissues lead to significant difficulties to achieve high diagnosis accuracy using hand-crafted features. Deep learning is an emerging machine learning technology that requires a relatively high computation power. Yet, it proved to be very effective in several difficult tasks that requires decision making at the level of human intelligence. In this paper, we develop a new network architecture inspired by the U-net structure that can be used for effective and early detection of breast cancer. Results indicate a high rate of sensitivity and specificity that indicate potential usefulness of the proposed approach in clinical use.