Detection of Her2 Scores and Magnification from Whole Slide Images in Multi-Task Convolutional Network

Abstract

For identifying Human epidermal growth factor receptor 2(HER2) Scores and magnification in different histopathology whole slide images(WSIs) of Breast cancer(BCa), an efficient multi-task convolutional Neural Network(CNN) is proposed. To prove its efficiency, we compare the accuracy, precision and recall with several mature Neural Network like, Lenet, AlexNet and Vgg16 from real data of our own, and show detection of the region of interest(ROI) on the heat map with the proposed multi-task CNN. Because of training model, we need to get mass data. Based on this, this paper also propose two new ways of data preparing with Label propagation(LPA) and SVM, which is used to extract enough typical patches automatically instead of tedious manual selection. After that, the data augmentation are used in the typical patches.