Breast Tumor Tissue Image Classification Using DIU-Net.

Area of Science:

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

Background:

• Pathologists focus on nuclei in pathology images.
• Existing models may lack interpretability and struggle with varying nuclei sizes and data imbalance.

Purpose of the Study:

• To develop a joint segmentation-classification model for breast cancer pathology images.
• To improve classification performance and model interpretability by mimicking pathologist's visual focus.

Main Methods:

• Proposed DIU-Net, a segmentation network with cross-scale description ability.
• Implemented Complementary Color Conversion Scheme to enhance generalization.
• Utilized dice loss and focal loss to address data imbalance.
• Adopted a joint training scheme for segmentation and classification networks.

Main Results:

• Achieved high binary/multi-class classification accuracy: 97.24%/93.75% (200×) and 98.19%/94.43% (400×) on the BreaKHis dataset.
• The model provides attention maps, increasing interpretability.
• Outperformed existing methods in classification performance.

Conclusions:

• The proposed joint segmentation-classification model significantly improves breast cancer pathology image analysis.
• DIU-Net offers enhanced accuracy, generalization, and interpretability, aiding pathologists in diagnosis.