Many medical imaging applications entail the acquisition of images from multiple points of view: in mammography two views are acquired in orthogonal directions which allow the radiologist, for instance, to rule out more easily false positives, which often appear only on a single view. This research thesis introduces and analyses a multi-view deep learning architecture that is trained for the task of breast cancer lesion detection on registered cranio-caudal (CC) and medio-lateral oblique (MLO) mammography views. The main innovation behind this methodology lies in the dual-stream structure of the network and a preliminary registration of the CC and MLO views that are jointly fed in the multi-view network as opposed to the usual one-at-a-time image processing of current object detectors, which is the dominant approach in the field.