Deep Multi-Magnification Networks for multi-class breast cancer image segmentation.

Area of Science:

• Digital pathology
• Artificial intelligence in oncology
• Computational imaging

Background:

• Pathologic analysis of breast carcinoma surgical specimens is crucial for treatment planning.
• Manual review of histologic slides is time-consuming and subjective.
• Digitized whole slide images offer potential for automated analysis.

Purpose of the Study:

• To develop an automated multi-class tissue segmentation method for breast cancer.
• To improve the efficiency and accuracy of surgical excision assessment.
• To present a novel deep learning architecture for analyzing whole slide images.

Main Methods:

• Developed a Deep Multi-Magnification Network (DMN) utilizing partial annotation.
• Employed a multi-encoder, multi-decoder, and multi-concatenation architecture.
• Trained and evaluated the network on patches from multiple magnifications of digitized whole slide images.

Main Results:

• The DMN achieved the highest mean intersection-over-union (IoU) compared to other architectures.
• Demonstrated superior performance in automated multi-class tissue segmentation.
• Outperformed both single and multi-magnification-based approaches.

Conclusions:

• The proposed DMN effectively automates tissue segmentation in breast cancer histopathology.
• This AI-driven approach can assist pathologists in evaluating surgical excision completeness.
• Facilitates more accurate and efficient breast cancer assessment.