Deep Multiple Instance Learning for Automatic Breast Cancer Assessment Using Digital Mammography

Breast cancer causes serious public health problems; it is the most common cancer among women worldwide. Screening and early detection of signs of breast cancer increase the chance of survival. Early diagnosis is a crucial task for radiologists and physicians. Therefore, many computer-aided detection and diagnosis (CADx) systems are being developed to ensure the survival of radiologists' decisions. In this article, we describe a framework to automate assessment of suspicious regions, detected in screening mammography, without having carried out additional examinations, especially unnecessary biopsies in the case where the suspect regions are benign tumors. The setup of the proposed framework is ordered as follows: regions of interest (ROIs) have been segmented using a modified K-means algorithm; the bidimensional empirical mode decomposition (BEMD) algorithm is applied to derive many layers [bidimensional intrinsic mode function (BIMF)] from ROIs. Then, textural features are extracted from the obtained ROIs. First, directly from segmented ROI, second from the ROI and its sublayers (BIMFs + Residue). The features extracted in the second time have been grouped into a bag descriptive of the ROI under consideration. This bag is the input parameter of the classification algorithm based on the support vector machine (which has been confirmed to be beneficial for the classification of breast cancer). The average obtained sensitivity, specificity, accuracy, and area under the receiver operating characteristic (ROC) curve (AUC) rates, are, respectively, 98.60%, 98.65%, 98.62%, and 98.23%. Generally, the experimental results in INbreast, digital database of screening mammography (DDSM), and Mammography Image Analysis Society (MIAS) datasets demonstrate the robustness and the efficiency of the developed framework compared to previous works in the literature and have shown a significant advance.