Malignant vs. Non-malignant Annotations on TCGA Breast Cancer Whole Slide Images for AI Analysis

Area of Science:

• Digital pathology
• Computational oncology
• Artificial intelligence in medicine

Background:

• Accurate identification of malignant and non-malignant regions in breast cancer whole slide images (WSIs) is crucial for understanding tumor heterogeneity and histopathological evaluation.
• Annotated data is essential for training deep learning models to analyze histological structures and features.

Purpose of the Study:

• To create a high-quality annotated dataset of breast cancer WSIs for AI model development and benchmarking.
• To assess the quality of manual annotations through inter-observer agreement and machine learning model performance.

Main Methods:

• Acquired 50 breast cancer WSIs from The Cancer Genome Atlas (TCGA).
• Expert pathologist manually annotated malignant (n=1,882) and non-malignant (n=374) regions using QuPath.
• Independent review by a second pathologist to ensure inter-observer agreement.
• Trained a hybrid contrastive-supervised machine learning model for patch-level classification to assess annotation quality.

Main Results:

• Achieved a 99.95% inter-observer agreement between pathologists.
• The machine learning model attained an F1-score of 0.90, indicating high-quality annotations.
• The developed dataset is suitable for benchmarking and developing AI models for breast cancer histopathology.

Conclusions:

• The study presents an expert-quality annotated dataset of breast cancer WSIs.
• The dataset serves as a valuable resource for advancing AI in breast cancer histopathology.
• High annotation quality was validated through robust inter-observer agreement and machine learning performance.