MTGA-Net A Multi-Tier Global-Attention Deep Learning Framework for Enhanced Breast Mass Classification in Mammographic Images

Accurate and early detection of breast masses in mammographic images is critical for effective diagnosis and treatment of breast cancer. This study presents MTGA-Net, a multi-tier global-attention deep learning framework designed to enhance the classification of breast masses as benign or malignant. The framework integrates two specialized models: HiRes-GANet (Hierarchical Residual–Global Attention Network) and Mammo-MLA (Mammogram Multi-resolution Lightweight Attention), both optimized to capture fine-grained and global features across multiple resolutions. HiRes-GANet employs hierarchical residual blocks combined with global attention mechanisms to emphasize diagnostically significant regions while maintaining deep feature representations. Mammo-MLA is a lightweight network that uses multi-resolution attention gates to efficiently extract informative features and combine coarse and fine details, providing a balance between accuracy and computational efficiency. The framework is trained and evaluated on the publicly available Breast Cancer Mammographic Dataset from Kaggle, comprising annotated mammograms with varying lesion sizes and appearances. Grad-CAM visualization is incorporated to provide interpretability, highlighting regions contributing to model predictions and enabling better clinical insight. Experimental results demonstrate that MTGA-Net improves classification performance and robustness compared to conventional approaches, while Grad-CAM facilitates understanding of model decisions. This framework can support radiologists in enhancing diagnostic accuracy and reducing misclassification, particularly in cases with subtle or complex mass patterns. Future work includes extending the approach to multi-view mammograms and other lesion types, as well as optimizing for deployment in clinical settings with limited computational resources.

Keywords: Breast mass classification, Mammography, Deep learning, Global attention, HiRes-GANet, Mammo-MLA, Grad-CAM, Multi-resolution feature extraction.