TransBreast-Net: An interpretable vision transformer ensemble for breast cancer histopathology classification.

Area of Science:

• Medical Imaging
• Computational Pathology
• Artificial Intelligence in Oncology

Background:

• Histopathology classification of breast cancer presents challenges due to complex tissue morphology, low inter-class variability, and time-consuming manual analysis.
• Current deep learning methods, while effective, often struggle with long-range spatial dependencies and lack clinical interpretability.

Purpose of the Study:

• To develop an interpretable vision transformer ensemble framework, TransBreast-Net, for accurate and robust breast cancer histopathology classification.
• To enhance model interpretability by identifying diagnostically relevant tissue regions.

Main Methods:

• Utilized transfer learning with extensive data augmentation and preprocessing on BreakHis and ICIAR datasets.
• Employed an ensemble of three transformer architectures (CaiT, DeiT, Swin) to capture local and global features.
• Implemented ensemble strategies (Swin + DeiT for binary, Swin + CaiT for multi-class) for improved robustness and generalization.
• Integrated Gradient-weighted Class Activation Mapping (Grad-CAM) for visual explanations.

Main Results:

• Achieved 99.35% accuracy in binary classification (benign vs. malignant).
• Attained 97% accuracy in multi-class classification (benign, in situ, invasive, normal).
• Demonstrated enhanced interpretability through visual explanations highlighting diagnostically relevant regions.

Conclusions:

• TransBreast-Net offers high classification accuracy and robustness for breast cancer histopathology.
• The framework's interpretability makes it a promising tool for clinical decision support in AI-driven cancer diagnosis.