Breast Cancer Histopathological Image Classification Based on Graph Assisted Global Reasoning

Breast cancer ranks as the most prevalent cancer among women globally. Histopathological image analysis stands as one of the most reliable methods for tumor detection. This study aims to utilize deep learning to extract histopathological features and automatically identify tumor information, thereby assisting doctors in high-precision pathological diagnosis. This study proposes a dual-stream global-local network (DSGLNet) for breast cancer histopathological image classification. The proposed DSGLNet employs a dual-stream feature extraction architecture that leverages a convolutional network to extract local image features and employs graph convolutional mapping to construct a global feature interaction space for capturing global information. By deeply integrating both local and global features, the network achieves precise image classification. Additionally, image preprocessing through feature engineering normalizes image colors and enhances the details of tissue cell boundaries. The proposed DSGLNet model was thoroughly evaluated on the publicly available BreakHis dataset, encompassing different magnification levels for the identification of tumor nature and tumor types. The 40 magnification histopathological images achieved the best diagnostic results, with the identification of tumor nature reaching 0.966 accuracy and 0.973 precision, outperforming other advanced methods.