Automatic Breast Cancer Grading of Histological Images using Dilated Residual Network

Abstract

Breast cancer is one of the leading causes of female death worldwide. Histological evaluation of the breast biopsies is essential in the early detection. Recently, deep learning methods are developed to automatically grade breast cancer of histological images. For the critical local and global features of histological images, few existing deep learning methods effectively extract both of them. Most methods extract one at the loss of the other, with degraded multi-class classification accuracy. In this paper, we propose an effective breast cancer classification method of histology images based on a modified dilated residual network (DRN). The proposed method effectively captures the global feature while maintaining the local information, and thus achieves notably high multi-class classification accuracy. Experimental results show that for the four-class breast cancer classification problem, an accuracy of 89.5% can be obtained, which outperforms all the prevalent methods. In comparison to the manual diagnosis accuracy of 89% from pathologists, the proposed automatic diagnosis method is practical and promising.