Predicting estrogen receptor status from HE-stained breast cancer slides using artificial intelligence

Area of Science:

• Computational pathology
• Digital pathology
• Artificial intelligence in oncology

Background:

• Estrogen receptor (ER) assessment is crucial for breast cancer patient stratification.
• Pathology labs face increasing workloads and high costs for molecular analyses.
• Hematoxylin and eosin (HE)-stained slides offer a potential alternative for predictive modeling.

Purpose of the Study:

• To develop an AI model for predicting ER status from HE-stained tissue microarrays (TMAs).
• To demonstrate the feasibility of predicting complex molecular analyses using digital pathology.
• To enhance efficiency and potentially reduce costs in pathology laboratories.

Main Methods:

• Utilized a clustering-constrained attention multiple-instance learning (CLAM) framework.
• Trained and tested models on TMAs from over 2,000 Norwegian breast cancer patients.
• Evaluated different patch sizes and CLAM configurations with extensive hyperparameter tuning.

Main Results:

• Achieved high performance metrics on internal and external test sets (e.g., AUC up to 0.95).
• Demonstrated superior classification performance with larger patch sizes.
• Validated the model's predictive capability on unseen data.

Conclusions:

• AI-based prediction of ER status from HE slides is feasible and accurate.
• This approach can serve as a proof-of-concept for predicting other molecular markers.
• Potential to streamline pathology workflows and reduce diagnostic costs.