Sliding window based deep ensemble system for breast cancer classification

Breast cancer is a severe problem for women around the world especially in developing countries, according to recent reports from the World Health Organization (WHO). High accuracy and early detection of breast cancer reduces the mortality rate, in the other hand, recognition of breast cancer is a complicated issue. Various studies and methods have been carried out to overcome this problem and to obtain accurate screening of breast cancer. One of the most recent methods with high performance is deep learning; it has been used to classify breast cancer using mammograms or histopathological images. This paper proposes a new using the concept of sliding window, and using the ensemble of four pre-trained convolutional neural networks (CNN) in order to classify breast cancer into eight classes. In this study, each image produces 4 non-overlapped sliding windows which are fed to GoogleNet, AlexNet, ResNet50, and DenseNet-201 CNNs, and an ensemble is then done to find the major class of each window, the ensemble is then applied again to find the class of the whole histopathological image. Breast Cancer Histopathological Database (BreakHis) database has been employed in this paper with eight classes (Adenosis, Ductal Carcinoma, Fibroadenoma, Lobular Carcinoma, Mucinous Carcinoma Papillary Carcinoma, Phyllodes Tumour, Tubular Adenoma). The proposed method is applied to four magnification cases: 40x, 100x, 200x, and 400x images. The proposed ensemble technique achieved an accuracy of 99.3325%. The results of the proposed system are comparable to recent studies results.