Breast cancer is a leading cause of mortality among women worldwide, highlighting the urgent need for early screening and accurate detection. Although ultrasound imaging is widely used for its safety and accessibility, the scarcity of expert radiologists calls for reliable automated solutions. This study addresses the challenge of tumor classification in limited data settings from breast ultrasound images. The proposed Classification with Auxiliary Representation Learning (CARL) framework is a unified, one-stage learning system that integrates self-supervised representation learning with lesion-aware classification. Unlike existing two-stage or purely supervised methods, CARL learns semantically rich features while focusing on clinically relevant tumor regions. The framework is encoder-agnostic and improves performance across both CNN and ViT architectures. Extensive experimentation demonstrates that the CARL framework outperforms state-of-the-art models by a notable margin across all metrics. Grad-CAM visualization corroborates that the model bases its prediction on actual tumor lesions, while improving clinical reliability and interpretability.