Skin cancer is one of the most common cancers globally, and early detection appreciably improves the affected person’s survival. Automated skin lesion analysis, using deep learning, suggests outstanding promise; however, performance is sensitive to picture preprocessing, segmentation, and feature representation. In this observation, we compare four preprocessing pipelines on the HAM10000 dataset to identify the simplest approach for lesion segmentation and classification. Segmentation is performed using a U-Net with sonar-inspired visual enhancement and morphological post-processing. Segmented lesions are classified into seven diagnostic categories using 25 pretrained models, including CNNs, Vision Transformers, and lightweight architectures. Segmentation is evaluated with Dice, Jaccard, sensitivity, specificity, and accuracy, while classification uses precision, recall, F1-score, and accuracy. Results show that advanced preprocessing substantially improves performance, providing a robust framework for automated skin lesion analysis and highlighting the critical role of preprocessing in deep learning-based medical imaging.