Cell Segmentation in Triple-Negative Breast Cancer Histopathological Images Using U-Net Architecture

Abstract

In this study, cell segmentation study is performed on histopathology dataset obtained from triple-negative breast cancer patients. The images present in the triple-negative breast cancer (TNBC) nuclei segmentation dataset are reproduced by the artificial data generation method. By using groundtruth images in the dataset, a second groundtruth image set containing only cell boundaries is created. The dataset, which was separated as training and test data, is passed through histogram equalization process and divided into patches in size of 64x64, 128x128 and 512x512 pixels. A total of three different datasets and two different groundtruths are evaluated separately and training is conducted with U-Net. The output image obtained as a result of the training is improved by thresholding. Output images obtained from 64x64 and 128x128 pixel sizes are combined with a technique similar to the ensemble learning method and converted into 512x512 pixel images. The accuracies of all steps performed during the study are presented in tables.