HMAT-Net: A compact hierarchical multi-scale attention network with feature tokenization for multi-class breast histopathological image classification

Breast cancer histopathological image classification is challenging due to considerable morphological variations across magnification levels, high visual similarity between tumour subtypes, class imbalance, and the limited ability of conventional convolutional neural networks to capture long-range contextual information. To address these challenges, this study presents HMAT-Net, a compact hierarchical CNN–attention framework for multi-class breast histopathological image classification. The architecture integrates a dual-attention convolutional stem, an Attention-Enhanced Multiscale (AEM) block for multi-scale representation learning, a Feature Tokenization Attention Block (FTAB) for global contextual modeling, and a CBAM-based refinement module followed by complementary global average and max pooling. HMAT-Net contains approximately 0.88 million trainable parameters while jointly modeling local, multi-scale, and global feature representations. On the BreakHis dataset, HMAT-Net achieves classification accuracies of 97.53%, 97.58%, 97.75%, and 97.87% at 40×, 100×, 200×, and 400×magnifications, respectively, with an overall multi-class accuracy and Macro-F1 score of 98.37%. Comparative evaluation against established deep learning baselines demonstrates that HMAT-Net achieves competitive or higher classification performance while maintaining a substantially smaller parameter footprint. Module-wise ablation experiments further examine the contribution of the proposed architectural components. Grad-CAM++ visualization further provides qualitative insight into the spatial regions influencing model predictions. Overall, the results indicate that HMAT-Net offers a compact and effective framework for multi-class breast histopathological image classification, with a favorable balance between classification performance and model complexity.