Classification of Breast Cancer Histopathology Images by Cell-Centered Deep Learning Approach

Abstract

Breast cancer is one of the most common cancer types worldwide today. The diagnosis of this cancer is usually made by the intensive work of pathologists on stained biopsy tissue images. In this study, breast cancer tissues are classified into four classes (normal, in situ, invasive and benign) by using the convolutional neural networks. In the training and test process performed on histopathological images, a cell-centered approach is followed instead of using the whole image. The results are examined separately for both image patches and the classification of the whole microscopic image. In addition, the effect of image patch sizes and cell neighborhood relationships on accuracy in different dimensions is investigated. As a result, 75% in four classes and 80% accuracy in cancer/non-cancer two-grade evaluation were achieved with the application which was trained with the training data of BACH dataset and tested with the test data of Bioimaging2015 dataset.