Normal breast tissue (NBT)-classifiers: advancing compartment classification in normal breast histology.

Area of Science:

• Computational pathology
• Digital histopathology
• Artificial intelligence in oncology

Background:

• Quantitative analysis of normal breast tissue (NBT) is crucial for early cancer detection but remains limited.
• Computational histopathology using digitized slides offers potential but lacks AI-based analysis for NBT.

Purpose of the Study:

• To develop and validate AI models for the classification of normal breast tissue.
• To establish robust analytical tools for NBT compartment analysis to aid in early breast cancer detection.

Main Methods:

• Curated 70 whole slide images (WSIs) of NBT with pathologist-guided annotations.
• Developed convolutional neural network (CNN)-based, patch-level classification models (NBT-Classifiers).
• Validated models across three external cohorts using different patch sizes (128x128µm and 256x256µm).

Main Results:

• NBT-Classifiers achieved high performance with AUCs of 0.98-1.00 across external cohorts.
• The models identified distinct features of normal tissue, differentiating them from precancerous and cancerous epithelium.
• Explainable AI techniques visualized learned features, and integration into a pipeline facilitated peri-lobular region analysis.

Conclusions:

• NBT-Classifiers provide accurate, compartment-specific analysis of normal breast tissue.
• These tools enhance understanding of NBT morphology, serving as references for identifying premalignant changes.
• The developed models support early breast cancer prevention strategies through improved quantitative analysis.