Breast Cancer Hormone Receptor Status Determination from H&E-Stained Biopsy Images Using Pixel-Level Classifiers

Background: Analysis of digital images of histopathological sections is increasing due to widespread adoption of fully digitised workflows and the greater availability of whole-slide scanners. Currently, hormone receptor status in breast carcinoma is assessed by pathologists scoring separate immunohistochemically stained sections. Methods: In this study, we employed pathologist-verified pixel-level annotations to train nested pixel classifiers capable of making case-level predictions of oestrogen receptor (ER), progesterone receptor (PR), and human epidermal growth factor receptor 2 (HER2) status directly from H&E-stained sections using biopsy cases alone. The model was evaluated on both an internal test set and an external international evaluation set from an institution in a different continent using different scanner hardware without the need for image normalisation. Results: In the internal test set, the models achieved AUCs of 0.8030, 0.7956 and 0.7488 for ER, PR, and HER2, respectively, with AUCs of 0.7008 and 0.7488 for ER and PR using an external cohort from an institution from which no cases were used for training. Conclusions: Our data highlight a potential strategy by which a pixel-based classifier, typically developed to quantify histological features within individual cases, could be used to make case/slide-level predictions but illustrate the challenges associated with this approach.