Correlating nuclear morphometric patterns with estrogen receptor status in breast cancer pathologic specimens

Area of Science:

• Computational pathology
• Digital pathology
• Machine learning in oncology

Background:

• Hormone receptor status is critical for breast cancer treatment decisions.
• Traditional methods for determining estrogen receptor (ER) status involve immunohistochemistry (IHC).
• There is a need for objective and efficient methods to assess ER status.

Purpose of the Study:

• To develop and validate a machine learning (ML) framework for predicting clinical estrogen receptor (ER) status from hematoxylin and eosin (H&E)-stained breast cancer tissue samples.
• To explore the relationship between nuclear morphology and ER pathway activation.
• To enhance the interpretability of deep learning models in digital pathology.

Main Methods:

• A machine learning pipeline was developed to segment nuclei from H&E images.
• Nuclear position, shape, and orientation descriptors were extracted.
• A deep neural network was trained to predict ER status based on these spatial features.
• The model was trained on 57 invasive ductal carcinoma (IDC) tissue cores and tested on 56 independent samples.

Main Results:

• The ML pipeline achieved an AUC ROC of 0.72 (95% CI: 0.55-0.89) in predicting ER status on an independent test set.
• The study demonstrated that machine-derived morphologic features correlate with hormone receptor pathway activation.
• The deep neural network learned spatial relationships between pre-defined biological features, enhancing model interpretability.

Conclusions:

• This pilot study successfully demonstrates the potential of a machine learning framework to predict ER status from H&E images.
• The approach offers an interpretable alternative to other deep learning methods in pathology.
• Future research should focus on correlating morphometric features with quantitative ER status and predicting response to hormonal therapy.