A Breast Image Classification Method Based on Attention Feature Fusion

As the number of breast MRI patients continues to rise, enhancing the accuracy and reliability of breast tumor dynamic contrast-enhanced magnetic resonance imaging (DCEMRI) benign-malignant cancer tasks is crucial. In this study, we introduce a novel local-global attention feature fusion network, building upon the framework of the local-global cross-attention fusion network model, and utilizing its foundation. This method alleviates the quadratic complexity of the self-attention mechanism. By focusing on different spatial and scale DCE-MRI information, important features from the outputs of SENet and Transformer networks are fully explored and utilized. Significant results are achieved on the BreastDM dataset, with the highest accuracy reaching 88.51% and an AUC value of 94.2%. This optimizes the breast tumor classification task.