Abstract 5295: Toward high throughput immune infiltrate analysis from H&E stained images

Abstract

Abstract Our appreciation of the immune system in breast cancer is rapidly evolving. New methods to quantify the composition and spatial distribution of the immune infiltrate from pathology images are urgently needed to develop better strategies to activate the anti-tumor immune response. Existing methods to quantify the immune infiltrate are tedious (manual counting) or rely on immunohistochemistry (IHC) staining to identify immune cells. While significant efforts are underway to develop machine learning tools to categorize cells based on morphologic features in hematoxylin and eosin (H&E) stained histopathology images, all approaches are limited by the scarcity of large, annotated ground truth training sets. In this work, we propose a new approach, leveraging “style-transfer” algorithms from the computer vision community to generate large quantities of training data from IHC stains. Using style transfer, we generate synthetic H&E images from IHC stains for immune markers, such as CD45, CD8 and CD4. The generated images have pixel-perfect ground truth provided by IHC, and do not need manual labeling. We investigate the impact of training on increasing numbers of synthesized images and validate our algorithms on an independent test set annotated by a pathologist (n=100 patients, 3064 identified lymphocytes). Our baseline classifier achieves a 83% accuracy, compared to the pathologist ground truth. This work provides a novel, tractable, and efficient way to train data-hungry algorithms to identify multiple cell types from H&E stained images. This work is supported by a grant from the Breast Cancer Research Foundation (BCRF-17-002) Citation Format: Rishi R. Rawat, Daniel Ruderman, David B. Agus. Toward high throughput immune infiltrate analysis from H&E stained images [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2018; 2018 Apr 14-18; Chicago, IL. Philadelphia (PA): AACR; Cancer Res 2018;78(13 Suppl):Abstract nr 5295.