Background: Estrogen receptor (ER) status is a key biomarker for guiding breast cancer prognosis and therapy, with ~80% of tumours classified as ER-positive (ER+). Although immunohistochemistry (IHC) remains the gold standard to determine ER status, IHC requires considerable time and resources. To address this, we developed a semi-supervised deep learning model using Clustering-constrained Attention Multiple Instance Learning (CLAM) to automatically predict ER status from hematoxylin and eosin (H&E) whole slide images (WSIs) and prioritize cases requiring subsequent IHC & molecular analysis.

Methods & Design: A total of 776 diverse H&E WSIs with corresponding ER labels were obtained from the TCGA Breast Invasive Carcinoma cohort. The WSIs were segmented and tiled into non-overlapping patches, then split into training (n = 636), validation (n = 70), and independent generalization test (n = 70) sets. Feature embeddings were extracted using the ResNet50 convolutional neural network, and a semi-supervised CLAM framework incorporating attention-based multiple instance learning was subsequently trained on labeled cases (ER-positive [ER+] vs. ER-negative [ER−]). Model performance was evaluated using 10-fold cross-validation and an independent generalization set with standard evaluation metrics.

Results: The model achieved a mean +- SD area under the receiver operating characteristic curve (AUROC) of 0.88 +- 0.03 (95% CI [0.86, 0.90]) during 10-fold cross-validation and an AUROC of 0.94 (95% CI [0.89, 0.99]) on the independent generalization test set, demonstrating strong discrimination between ER+ and ER− cases, with a sensitivity of 0.90, specificity of 0.94, and balanced accuracy of 0.92.

Conclusion: A semi-supervised CLAM model applied to H&E WSIs accurately stratifies ER status and may support IHC triage, potentially reducing diagnostic turnaround time and ensuring timely treatment for patients most likely to benefit from IHC and molecular analysis.

Learning Objectives:

1. The audience will be able to describe how deep learning models predict ER status from H&E slides.
2. The audience will be able to interpret ROC–AUC and key metrics in ER prediction models, and evaluate ER stratification
3. Understand the importance of AI-based ER assessment in triaging breast cancer to improve workflow and turnaround time & care.