A Hybrid Convolutional and Recurrent Deep Neural Network for Breast Cancer Pathological Image Classification

Hematoxylin and Eosin (H&E) stained breast tissue samples from biopsies are observed under microscopy for the gold standard diagnosis of breast cancer. However, a substantial workload increases and the complexity of the pathological images make this task time-consuming and may suffer from pathologist's subjectivity. Facing this problem, the development of automatic and precise diagnosis methods is challenging but also essential for the field. In this paper, we propose a new hybrid convolutional and recurrent deep neural network for breast cancer pathological image classification. Our method considers the short-term as well as the long-term spatial correlations between patches through RNN which is directly incorporated on top of a CNN feature extractor. Experimental results showed that our method obtained an average accuracy of 90.5% for 4-class classification task, which outperforms the state-of-the-art method. At the same time, we release a bigger dataset with 1568 breast cancer pathological images to the scientific community, which are now publicly available from http://ear.ict.ac.cn/?page_id=1576. In particular, our dataset covers as many different subclasses spanning different age groups as possible, thus alleviating the problem of relatively low classification accuracy of benign.