HER2 classification in breast cancer cells: A new explainable machine learning application for immunohistochemistry

Area of Science:

• Biomedical Engineering
• Computational Pathology
• Oncology

Background:

• Immunohistochemistry (IHC) for HER2 in breast cancer diagnosis has qualitative limitations and high inter-observer variability.
• Complementary techniques like fluorescent in situ hybridization (FISH) are often required to confirm diagnosis.
• Automated algorithms are needed to improve the accuracy and consistency of HER2 biomarker classification for patient treatment.

Purpose of the Study:

• To demonstrate that an explainable Machine Learning (ML) model can improve the diagnostic value of HER2 IHC analysis.
• To identify criteria that enhance IHC interpretation using ML and SHAP values.
• To optimize HER2 classification by integrating final diagnostic data (IHC + FISH) into the ML model training.

Main Methods:

• A supervised ML model (logistic regression) was trained on 393 HER2 IHC microscopy images from 131 breast cancer patients.
• The model used both IHC-only and IHC + FISH diagnoses as training outputs.
• Explainability was achieved using Shapley Additive exPlanations (SHAP) values to interpret model decisions.

Main Results:

• The ML model achieved better discrimination between amplified and normal HER2 expression when trained with IHC + FISH final diagnoses (AUC 0.94) compared to IHC-only diagnoses (AUC 0.81).
• SHAP analysis indicated that membrane distribution patterns were more analytically impactful than signal intensity.
• Model performance improved by downplaying signal intensity and emphasizing subcellular staining patterns when trained with definitive diagnoses.

Conclusions:

• An explainable ML model can enhance HER2 IHC analysis for breast cancer diagnosis.
• Focusing on subcellular staining patterns, rather than just signal intensity, improves diagnostic accuracy.
• This approach offers a pathway to refine pathological diagnosis prior to FISH consultation, reducing variability and improving patient management.