Invasive Region Segmentation using Pre-trained UNet and Prognosis Analysis of Breast Cancer based on Tumor-Stroma Ratio

Abstract

Breast cancer (BCa) is a type of disease that has multiple prognostic markers that differs from one cancer stage to another. Assessing the area and pattern of cancer regions is essential for pathological investigations. However, the main purpose of this study is to segment the regions of invasive carcinoma (i.e., non-tubular and tubular) in the histological sections of BCa. The segmentation was performed on hematoxylin and eosin (H&E)-stained tissue slides of 42 BCa patients from 5 different centers in Korea. The tumor-stroma ratio (TSR) is a promising prognostic parameter for BCa as well as in other epithelial cancer types estimating the area of stroma and tumor regions. Therefore, in this paper, we used the pre-trained convolutional neural network (CNN) models as a backbone of UNet, to precisely extract the tumor regions from the stroma tissue components for TSR analysis.