MS-RepViT: A Multi-Scale Feature Enhanced Reparameterized Vision Transformer for Breast Cancer Histopathology Image Classification

As an important threat to the health of women around the world, the accurate detection of breast cancer pathological images is of critical significance for early diagnosis and treatment. In recent years, deep learning technology has made significant progress in medical imaging analysis, but traditional models still have shortcomings in multi-scale capture and microstructure recognition. This paper proposes an improved MSRepViT network, with MS Block (multi-scale convolutional block), scSE (Spatial and Channel Squeeze and Excitation) and GAM (Global Attention Mechanism), to enhance the capacity to extract microstructure and global semantic information of breast cancer pathological images. Specifically, MS Block expands the respective field through the parallel structure of heterogeneous convolution kernels and improves multi-scale feature expression. And the fusion attention mechanism effectively alleviates the difficulties in feature conflicts and information fusion, and significantly improves the classification performance of the model. The experiments verified the superiority of the proposed model on the disclosed BreakHis dataset and achieved a classification accuracy of 97.12 % to 97.74 % at different magnifications, which was better than a variety of mainstream deep learning models. At the same time, the model parameters and calculation amount were effectively controlled. The results show that this method has good application potential in automatic classification of breast cancer pathological images, and provides strong technical support for clinical auxiliary diagnosis.