Early detection of breast cancer has become an essential medical procedure to reduce mortality rates in patients. Although various deep learning methods enhance cancer analysis and detection, including feature extraction techniques, transfer learning, image fusion, and data augmentation, challenges such as limited data and class imbalance impede progress in early cancer detection. This study introduces a new approach that employs a dual ensemble learning and MixUp data techniques to simultaneously (1) tackle unbalanced classes and (2) improve diversity in ultrasound images. The proposed training methodology involves various strategies, integrating original and augmented datasets to maintain model training and increase performance. Through comprehensive evaluations using benchmark breast ultrasound images (BUSI), the proposed approach demonstrates a substantial improvement and feasibility in the multi-classification of breast cancer.