Improved Multi-Category Classification of Breast Cancer Histopathology Images Using Weighted Cross-Entropy Loss and Convolutional Block Attention Module

In this study, we proposed an improved approach for multi-category classification of breast cancer histopathological images using the BreakHis dataset, which contains significant category imbalance. To address this issue, we applied a Weighted Cross-Entropy (WCE) loss function, which gave more weight to underrepresented categories, thereby enhancing the model’s ability to classify them correctly. Given the complex and heterogeneous nature of pathological tissue images, we also integrated the Convolutional Block Attention Module (CBAM) to enable the model to focus on the most relevant features. This selective attention mechanism, inspired by human expert diagnosis, improved performance in distinguishing subtle tissue pattern differences. Our experiments showed that combining WCE and CBAM significantly outperformed baseline models across various magnification levels, with substantial improvements in F1-score, recall, and precision. These results demonstrated the effectiveness of our method in addressing both category imbalance and the inherent complexity of pathological image classification. Finally, we discussed the limitations of our approach, potential future directions, and its relevance to realworld clinical applications.