Abstract
Breast cancer is the most widespread cancer amongst women worldwide for cancer diagnoses in the world. In the first place, screening mammography is widely acknowledged to be the most effective imaging tool which allows to detect breast cancer early. Early detection is particularly associated with a decrease of incidence and mortality rate. However the limiting factors with the use of mammography are i) the fact that mammogram is difficult to interpret due to the high density of breast tissue, ii) the workload of radiologists; the situation is even worse because of the double reading process, and iii) false positive recalls are many times accompanying with needless tests and biopsies. We propose to set up a multi-view based design of a deep convotutional neural network in order to carry out mammography screening task - the fulfilled network can extract distinctive characteristics from Medio-Lateral Oblique (MLO) and Cranial Caudal (CC) mammography's views for each breast (a set of 4 images). We test it on a subset selected from the open Digital Database for Screening Mammography (DDSM) using exams (each exam is composed of 4 images). We demonstrate that our approach can outperforms existing ones both in the matter of prediction accuracy and false positive rate reduction. Our method accomplishes a 97.77% specificity rate and a 98.7% accuracy rate.