Contrasting low- and high-resolution features for HER2 scoring using deep learning.

Area of Science:

• Oncology
• Computational Pathology
• Biomarker Analysis

Background:

• Accurate HER2 (Human Epidermal growth factor Receptor 2) classification in breast cancer is crucial for targeted therapy selection.
• Traditional immunohistochemistry (IHC) classification is subjective and suffers from inter-observer variability.
• A 3-way classification (0, low, high) is essential for identifying patients who may benefit from HER2-targeted therapies.

Purpose of the Study:

• To develop and evaluate deep learning models for automated 3-way HER2 classification using IHC images.
• To introduce the India Pathology Breast Cancer Dataset for research in automated breast cancer diagnosis.
• To compare the performance of an end-to-end deep learning model against traditional patch-based methods.

Main Methods:

• Development of the India Pathology Breast Cancer Dataset with HER2 IHC slides from 500 patients.
• Training and evaluation of various deep learning models, including an end-to-end ConvNeXt network.
• Utilized low-resolution IHC images for the ConvNeXt model to assess classification performance.

Main Results:

• The end-to-end ConvNeXt network achieved an overall F1 score of 83.52% for 3-way HER2 classification.
• This represents a 5.35% improvement over patch-based methods.
• Class-wise F1 scores were 75.6% (HER2-0), 82.4% (HER2-low), and 91.5% (HER2-high), with challenges noted in differentiating HER2-0 and HER2-low.

Conclusions:

• Deep learning techniques, particularly the end-to-end ConvNeXt model, show significant potential for accurate and reproducible HER2 classification in breast cancer.
• Automating HER2 classification can reduce pathologist workload and inter-observer variability.
• Integration of these AI tools into clinical workflows may enhance patient outcomes through optimized targeted therapy selection.