Classification of H&E Stained Breast Cancer Histopathology Images Based on Convolutional Neural Network

Abstract

Automatic classification of H&E breast cancer histopathology images is a challenging task. Computer-aided diagnostic systems help reduce costs and increase the efficiency of the process. Although the existing research on breast cancer histopathology image classification is higher than 90% accurate in the binary classifications (non-carcinoma/carcinoma), the classification accuracy of four classifications (normal, benign, in situ, invasive) is less than 80%. This paper proposes a framework for the classification of H&E stained breast cancer histopathological images, which includes two methods based on convolutional neural network. The first method is based on the convolutional neural network structure of the SE-ResNet module, and the second method is based on the transfer learning hybrid model structure, which achieves the accuracy of 80.33% and 86.11% respectively. Compared with the state-of-the-art method, the accuracy is improved by 2.56% and 8.33% respectively. The proposed framework achieves 91.67% accuracy in binary classification and is competitive with state-of-the-art methods.