Embedding-driven dual-branch approach for accurate breast tumor cellularity classification.

Area of Science:

• Medical Image Analysis
• Computational Pathology
• Artificial Intelligence in Oncology

Background:

• Accurate breast tumor cellularity classification is crucial for effective cancer treatment and prognosis.
• Histopathological images are key for evaluating cellularity, but manual assessment can be subjective and time-consuming.

Purpose of the Study:

• To develop and validate a novel dual-branch framework for precise breast tumor cellularity classification using histopathological images.
• To investigate the contribution of embedding-driven and vision-based approaches in cellularity classification.
• To enhance feature extraction and mitigate overfitting through a Knowledge Block.

Main Methods:

• A dual-branch framework integrating an Embedding Extraction Branch (Virchow2 transformation) and a Vision Classification Branch (Nomic AI Embedded Vision v1.5).
• Combination of outputs from both branches for final classification.
• Implementation of a Knowledge Block with fully connected layers, batch normalization, and dropout.
• Utilized data augmentation techniques to improve model performance.

Main Results:

• The proposed framework achieved high performance metrics, including accuracy, specificity, sensitivity, precision, and F1 score.
• Ablation studies confirmed the essential role of the embedding extraction branch, with its removal significantly decreasing accuracy.
• The vision classification branch also contributed significantly, with its removal leading to a smaller accuracy decrease.
• Data augmentation was shown to enhance model performance.

Conclusions:

• The dual-branch framework offers a robust and accurate method for breast tumor cellularity classification.
• Both embedding extraction and vision classification components are vital for optimal performance.
• The study highlights the potential of AI-driven approaches in digital pathology for improved cancer diagnosis.