Skin cancer is one of the most common types of cancer, and its early detection and treatment are crucial in the survival of those afflicted. Technological advancements have enabled the surge of fields such as Computed-Aided Diagnosis (CAD), which aims to help in disease assessment using tools such as convolutional neural networks. However, one of the main problems in medical image analysis is the lack of data, leading to imbalanced datasets that make generalization more difficult. In this work, a comparative analysis of classification and processing performance for skin cancer using the HAM10000 dataset on different data preprocessing approaches is conducted. The VGG19, ResNet50, EfficientNetV2, and MobileNetV3 models are compared based on the implementation of data augmentation, image segmentation, and generative data methods. The results showcase that the best performing models are ResNet50 and MobileNetV3 using data augmentation, achieving accuracy, F1 score and latency of 0.7067, 0.7093, 0.000207 for ResNet50, and 0.74, 0.7423, 0.000206 for MobileNetV3, respectively, demonstrating a positive classification and latency trade-off. Although image segmentation and generative data can lead to better model generalization, further experimentation is needed to achieve improvement over regular data augmentation.