MSWA-ResNet: Multi-Scale Wavelet Attention for Patient-Level and Interpretable Breast Cancer Histopathology...

Area of Science:

• Computational pathology
• Medical image analysis
• Deep learning in oncology

Background:

• Manual breast cancer histopathological classification is laborious and inconsistent.
• Existing deep learning methods may suffer from data leakage and lack multi-scale frequency analysis.

Purpose of the Study:

• To develop and evaluate MSWA-ResNet, a novel deep learning model for automated breast cancer classification.
• To enhance diagnostic accuracy and interpretability in histopathological analysis.

Main Methods:

• Implementing a Multi-Scale Wavelet Attention Residual Network (MSWA-ResNet) with recursive discrete wavelet decomposition.
• Utilizing a strict patient-level protocol on the BreakHis dataset with 70/30 patient-wise splitting and cross-validation.
• Employing ensemble prediction and hierarchical aggregation from patch to patient level.

Main Results:

• MSWA-ResNet achieved 96% patient-level accuracy at 100×, 200×, and 400× magnifications, and 92% at 40×.
• Demonstrated improved accuracy and F1-scores over baseline CNNs, with competitive parameter counts and inference times.
• Grad-CAM visualization confirmed improved localization of diagnostically relevant regions.

Conclusions:

• MSWA-ResNet effectively integrates multi-scale frequency information for accurate breast cancer classification.
• The proposed model offers a robust and interpretable solution for automated histopathological analysis.
• This approach addresses limitations of existing methods, paving the way for improved diagnostic workflows.