Interpretable HER2 scoring by evaluating clinical guidelines through a weakly supervised, constrained deep learning...

Area of Science:

• Digital Pathology
• Artificial Intelligence in Medicine
• Oncology Biomarkers

Background:

• Accurate Human Epidermal growth factor Receptor-2 (HER2) expression evaluation is crucial for breast cancer treatment selection.
• HER2 scoring exhibits high interobserver variability due to staining inconsistencies and visual estimation challenges.
• Existing methods lack interpretability for pathologists, hindering clinical adoption.

Purpose of the Study:

• To develop a semi-automatic, interpretable deep learning approach for HER2 scoring.
• To align AI-driven HER2 classification with American Society of Clinical Oncology/College of American Pathologists (ASCO/CAP) guidelines.
• To reduce interobserver variability in HER2 assessment.

Main Methods:

• A two-stage deep learning model was developed: tumor segmentation within a Region of Interest (ROI) followed by HER2 class classification.
• Weakly supervised, constrained optimization was used for classification, ensuring tumor surface percentage adherence to guidelines.
• A multi-pathologist consensus labeling strategy and supervised refinement of model outputs were employed.

Main Results:

• The model achieved an F1-score of 0.78 on the test set.
• The deep learning approach demonstrated interpretability for pathologists, providing HER2 class percentages.
• The system aids in assessing doubtful cases where pathologist consensus was not reached.

Conclusions:

• The proposed semi-automatic deep learning method offers an interpretable and accurate solution for HER2 scoring in breast cancer.
• This approach has the potential to standardize HER2 assessment and improve treatment selection.
• The study contributes to the development of interpretable AI models in digital pathology.