RI-ViT: A Multi-Scale Hybrid Method Based on Vision Transformer for Breast Cancer Detection in Histopathological Images

Breast cancer is one of the most significant health threats to women worldwide. This disease manifests through abnormal proliferation of cells and the formation of tumors in breast tissue. Definitive breast cancer diagnosis is usually determined by analyzing tissue samples obtained from biopsies and reviewing them by pathologists. However, this method is highly dependent on the knowledge and experience of pathologists and may lead to errors due to the subjective nature of human interpretation and the high volume of cases. This study presents a multi-scale hybrid model based on Vision Transformer and residual networks for breast cancer detection in histopathological images, abbreviated as RI-ViT. In this approach, local features are extracted through a combination of residual stages and multi-scale learning, while global features are obtained using the attention mechanism in transformers. This combination enables simultaneous extraction of both local and global features from histopathological images, effectively improving the model’s performance in detecting complex cases. We have used an imbalanced and publicly available dataset called BreakHis to evaluate the performance of the RI-ViT model. The experimental results of the proposed model show that it achieves accuracies of 99.75%, 98.80%, 98.01%, and 97.53% at magnifications of 40X, 100X, 200X, and 400X, respectively. The RI-ViT model can also perform well in an magnification-independent mode. Results show that, regardless of the magnification level, it achieves an accuracy of 99.37%, demonstrating its superiority over other state-of-the-art models.