Medical image diagnosis should take into consideration the information contained in multiple images, not just a single image, such as natural image classification. Mammography is the most basic X-ray screening method for diagnosing breast cancer, and mammograms have four images per patient. Convolutional neural networks (CNN) should be able to diagnose using these four images. This paper proposes a 4-channel input CNN that simultaneously concatenates four images to solve the multi-view problem. CNN using the proposed 4-channel input has been trained and validated with the digital database for screening mammography (DDSM). The network has shown an area under the ROC curve (AUC) of 0.952 for the 2-class (positive vs negative) problem. In addition, this thesis proposes a new approach for the localization of lesions without patch labels or mask labels.
