Comprehensive Experiments on Breast Cancer Hematoxylin and Eosin-stained Images Using UNet

The current challenge in breast cancer segmentation on hematoxylin and eosin-stained images lies in accurately capturing histologic phenotypes associated with cancer biomarkers. Many researchers have studied histologic patterns of cancerous regions of breast cancer images for a better understanding of diagnosis and treatment, yet it is not fully investigated because of its variability, complexity, and large data volumes. Comprehensive experiments on breast cancer segmentation can address this challenge by identifying heterogeneous cell regions in the tumor microenvironment and investigating a new methodology of segmenting histologic images by using advanced deep learning architectures. In this paper, we present findings in three experiments on breast cancer segmentation, exploring the effectiveness of a convolutional neural network called UNet identifying multiple heterogeneous cells such as blood, blood vessels, fat, glandular secretions, necrosis, and plasma cells in a tumor microenvironment, investigating the performance of UNet architecture in different sizes of cancerous regions such as a tumor, and tumor-infiltrating lymphocytes, and proposing a new methodology of a neural-style guided data augmentation focusing on the image segmentation of breast tumor-related nuclei in hematoxylin and eosin-stained images. The experiment results show that a modified UNet performs well on fat identification. Another experiment shows that the smaller input size of the tumor, stroma, and tumor-infiltrating lymphocyte provides better performance than the larger input size. Moreover, we demonstrate that the proposed method outperforms the traditional deep neural network models.