Deep Learning-Based Multi-Class Classification of Breast Digital Pathology Images.

Area of Science:

• Digital Pathology
• Artificial Intelligence
• Oncology

Background:

• Breast cancer is a major global health concern for females.
• AI systems are increasingly used for breast cancer image classification.
• Existing AI models often perform only binary classification, limiting clinical utility.

Purpose of the Study:

• To develop a clinically practical, AI-based multi-class classification system for breast digital pathology images.
• To advance beyond binary classification for more nuanced breast cancer diagnosis.

Main Methods:

• A two-stage architecture combining deep learning and machine learning methods was employed.
• The system was trained for multi-class classification: normal tissue, benign lesion, ductal carcinoma in situ, and invasive carcinoma.
• Model performance was evaluated on internal datasets and validated using public datasets (BreakHis, BACH).

Main Results:

• The AI system achieved 90.43% accuracy on test data at the Whole Slide Image (WSI)-level.
• Comparable accuracies were obtained on public datasets, demonstrating generalizability.
• High accuracy was also achieved on frozen section images for both multi-class (85.19%) and binary (96.30%) classification.

Conclusions:

• The proposed two-stage AI model effectively performs multi-class classification of breast pathology images.
• This AI system serves as a valuable tool to assist pathologists in breast cancer diagnosis.
• The model demonstrates strong generalizability across different datasets and classification tasks.