We'll describe our efforts in building an efficient convolutional neural network capable of automating breast cancer screening. First, we'll highlight fundamental differences between natural and medical images, as well as differences in current practices when training neural networks on these types of data. Second, we'll describe the architecture of our network, its training process and promising experimental results. Then we demonstrate how decisions of our network can be explained by visualizing parts of image that had the greatest influence on the predictions made. Our visualization reveals surprising agreement between radiologists and the network in spotting important regions of interest. Finally, we'll discuss future directions of research necessary to automate early diagnosis of breast cancer and beyond using medical imaging technology.