Multitask Classification of Breast Cancer Pathological Images Using SE-DenseNet

Abstract

Breast cancer has always been the main killer of women. The constantly development of Convolutional Neural Network (CNN) greatly improved the possibility of early diagnostics of breast cancer owing to its high efficiency and accuracy. In this paper, we apply the architecture of Densely Connected Convolutional Network (DenseNet), and then assimilate into Squeeze-and-Excitation Network (SeNet) to perform multitask classification on Camelyon16 which is a set of images of hematoxylin and eosin (H&E) stained breast histology microscopy. Whole-slide images (WSIs) are generally stored in a multi-resolution pyramid, our dataset contains patches of Camelyon16 under ×5, ×20, ×40 three magnifications. Our multitask is to identify the magnification of the patch and distinguish whether the extracted patch belongs to metastatic tumor area of WSIs at the same time by link two classifiers at the end of the same network. Whether on multitask or a single subtask, our network has showed excellent performance, SE-DenseNet-40 has even achieved an accuracy of 92.92% on CIFAR-10.