Fully automatic HER2 tissue segmentation for interpretable HER2 scoring.

Area of Science:

• Oncology
• Biomedical Image Analysis
• Computational Pathology

Background:

• HER2 overexpression is critical in breast cancer, guiding treatment decisions.
• Current HER2 scoring methods by human raters exhibit variability.
• Accurate HER2 assessment is vital for effective breast cancer therapy.

Purpose of the Study:

• To develop a fully automated and interpretable HER2 scoring pipeline for breast cancer.
• To align automated scoring with established clinical guidelines using pixel-level segmentation.
• To address the variability in human HER2 assessments through computational methods.

Main Methods:

• A fully automated pipeline utilizing pixel-level semantic segmentations for HER2 scoring.
• Employing polygon annotations to balance detail capture and annotation effort.
• Incorporating Wasserstein Dice loss for enhanced segmentation and a novel calibration step for scoring rules.

Main Results:

• The proposed pipeline achieves an F1 score of 0.832 for HER2 scoring.
• Demonstrated effectiveness in robust segmentation and accurate HER2 scoring.
• The calibration step positively impacts accuracy and consistency of automated scoring.

Conclusions:

• The developed pipeline offers a potent and reliable method for automated HER2 scoring.
• This approach can enhance the analysis of HER2 expression in breast cancer tissues.
• The automated system shows promise in reducing inter-observer variability in HER2 assessment.