Accurate automated classification of skin lesions remains a challenging problem due to subtle inter-class variations, substantial class imbalance, and variability in dermoscopic imaging conditions. In this study, we propose an attention-enhanced deep learning framework that integrates class-weighted loss and a targeted augmentation strategy to improve diagnostic accuracy and robustness. The self-attention mechanism enables effective modeling of global and local lesion characteristics, while class-weighted optimization mitigates imbalance-driven performance degradation. Comprehensive experiments, including ablation studies and benchmarking against the current state-of-the-art (IRv2 5 \(\times \) 5 + SA), demonstrate the effectiveness of the proposed approach. Our approach surpasses the state-of-the-art baseline, delivering improved overall classification performance and a pronounced boost in minority-class recall (0.886 vs 0.613) on the ISIC-2017 dataset. Augmentation analysis further confirms that rotation operation provide the most consistent performance gains, with F1-scores reaching 0.89. These results highlight the robustness, stability, and diagnostic utility of the proposed framework for skin lesion classification.