Abstract 5404: Predicting protein receptor status from H&E-stained images in breast cancer

Abstract

Abstract Correct prediction of cancer molecular subtype plays a very important role in determining the treatment of cancer patients. In determining the molecular subtype of breast cancer, protein receptors such as estrogen receptors (ER), progesterone receptors (PR), and human epidermal growth factor receptor type 2 (HER2) are used as key factors. These protein receptors also play an important role in determining the treatment method or predicting the prognosis of breast cancer. To confirm this, a test should be performed by immunohistochemical (IHC) staining. In this study, we investigated the morphological relationship between the molecular subtypes and hematoxylin and eosin (H&E) stained whole slide images (WSIs) without IHC stained images, and performed an experiment to evaluate that the protein receptors can be predicted by these morphological features. The TCGA-BRCA data were utilized in this study. There were 728 cases out of the total 1097 cases where the IHC status for ER, PR, and HER2 was either positive or negative. Each individual case is scanned using different scanners and magnifications. In some cases with multiple slides, we use only the first scanned image. The entire WSI was randomly split into 3:1:1, and used for training, tuning, and test, respectively. Individual WSIs was tiled into 1024 × 1024 patches for this study. The multi-task deep learning model predicts the protein receptor status of individual patches. In this study, a confidence measure was added to remove the uncertainty of the deep learning model. When learning WSIs, patches of less than 70% based on these confidence scores did not affect WSIs. Specifically, when learning morphological features, we used strong augmentation such as grayscale, gaussian blur, color jitter, and posterization to ensure that predictions were not biased by color alone. Based on individual patches, ER accuracy of 76%, PR accuracy of 65%, and HER2 accuracy of 79% are shown. As a result of prediction by majority voting on WSI images through patch predicted results, ER accuracy of 74.6% PR accuracy of 66%, and HER2 accuracy of 76.6% Through this study, it was confirmed that the sufficiently trained deep learning model predicted ER, PR, and HER2, which are important factors of molecular subtype, relatively well. Through this, it was found that there was some correlation between morphological characteristics and molecular subtypes in H&E stained WSIs. If more data are collected through future experiments, a molecular subtype prediction model can be developed. Citation Format: Geongyu Lee, Chungyeul Kim, Tae-Yeong Kwak, Sun Woo Kim, Hyeyoon Chang. Predicting protein receptor status from H&E-stained images in breast cancer. [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2023; Part 1 (Regular and Invited Abstracts); 2023 Apr 14-19; Orlando, FL. Philadelphia (PA): AACR; Cancer Res 2023;83(7_Suppl):Abstract nr 5404.