Despite the recent successes of screening mammography, breast cancer remains the second leading cause of cancer death in women and the leading cause of death in women aged 45 to 55. This suggests that there is an urgent need to improve the current approaches to breast cancer screening based on mammography. The specificity of the screening mammography is also limited as the continued large fraction of false-positives often leads to unnecessary follow-up procedures including biopsy that cause considerable pain, anxiety and cost. It is therefore necessary to improve diagnostic accuracy of mammography in terms of both sensitivity for early detection of breast cancer and specificity to keep the recall rate 10% or less as recommended by American College of Radiology. In this study, we investigate deep convolutional networks (DCN) for medical imaging by developing from scratch a novel family of DCNs based on the characteristics of medical images and evaluating it on screening mammographic exams. We design a novel, multi-view DCN that will work with a set of multiple views in their original resolution so as to avoid missing crucial micro-structures, opposed to using heavily downscaled images as often done with natural images or by some earlier work on applying DCN to medical images. Unlike most existing works which verify their approaches on a small number of cases, we propose to train and test the proposed networks on a large-scale data consisting of about 500,000 cases collected over the period of five years at our institution. Our preliminary experiments using an initial version of the multi-view deep convolutional network with about 25,000 cases strongly support the need for the proposed large-scale, highresolution deep learning for early-stage breast cancer screening. We have also collected 201,000 cases of screening mammograms to date and are in the process of annotating the cases based on the available followup studies. The overarching goal of this study is to develop, optimize and test a multi-view DCN system that can either independently read a series of screening mammographic exams to identify early breast cancers accurately or to assist a human reader so as to reduce reading time and improve confidence in early breast cancer detection. Aim 1 is to develop and optimize a multi-view DCN for automatic breast screening. Aim 2 is to conduct preliminary reader study on the impact of the MV-DCN as a screening exam aid system. The outcome of this study, the multi-view DCN for automatic breast screening and the preliminary assessment of its impact on human readers, will form an important basis on which further investigation into optimizing deep convolution networks for medical imaging can be conducted.