Breast cancer is a major global health concern, making early detection crucial for improving survival rates. Deep learning methods in this field face challenges such as class imbalance, varying imaging types, and clinical interpretability. This study introduces a lightweight transformer-based model, EFormer-EA, aimed at breast cancer classification across different imaging modalities. Our hybrid architecture uses EfficientFormerV2 for local feature extraction and External Attention for modeling global dependencies. We applied it to two public datasets: BreakHis, containing 7,909 histopathology images at four magnifications, and BUSI, with 830 ultrasound images across three classes. Preprocessing included modality-specific normalization, histogram equalization, and GPU-accelerated augmentation, with a class-weighted loss to address class imbalance. The EFormer-EA model achieved impressive results: an F1 score of 98.27% and a Matthew's correlation coefficient of 96.15 on BreakHis, and an F1 score of 98.46% and a PR-AUC of 99.43 on BUSI, outperforming existing models. We also incorporated Grad-CAM into a web application for real-time, explainable diagnosis. However, the study has limitations, including sensitivity to external memory sizes and dependence on just two datasets. Future work will focus on multi-center validation, edge deployment, and integration with federated learning.