Deep learning-based risk stratification for HER2-negative breast cancer patients

Abstract

In this paper, we present our analysis of the tumor microenvironment in digital pathology images to stratify risk in HER2-negative breast cancer patients using clinicopathological, spatial image, and cell-based features in a Coxs proportional hazard model. We start the analysis by processing a set of 304 training pathology images using our in-house pan-cancer trained tumor, stroma, and lymphocyte region identification convolutional neural networks. The next step is computation of spatial regions of interest, namely: lymphocytes within (and adjacent to) tumor, lymphocytes within (and adjacent to) stroma, and stroma within (and adjacent to) tumor areas. Various cell-level information in these regions are then summarized, augmented to clinicopathological data, and linked to patients survival to train a Coxs proportional hazards model. The proposed model outperformed a baseline model based on clinicopathological features only in analysis of an untouched test set of 202 whole slide images with p 8.49E-08 (HR = 0.4273).