Accurate and robust classification of skin lesions remains a critical challenge in dermatological AI due to issues such as visual similarity between lesion types and dataset imbalances. In this work, we propose a comprehensive framework to improve skin lesion classification by integrating three key strategies: lesion segmentation, synthetic mole collision simulation, and hierarchical learning. First, lesion segmentation is used to localize the mole and focus the model on relevant regions, reducing background noise. Second, we introduce a novel synthetic data generation technique that simulates mole collisions by combining two lesions into a single image, improving the accuracy of the model in case of multi lesions appearance. These synthetic images also serve as a form of data augmentation, enhancing model generalization. Finally, we employ hierarchical learning that predicts lesion classes and sub-classes. Experimental results demonstrate that while our multi-label model may be slightly outperformed on class-specific metrics such as sensitivity, it achieves superior or comparable performance on global metrics like AUROC and specificity. This study highlights the potential of combining structural priors, synthetic augmentation, and label hierarchy to advance robust skin lesion classification.