Deep Learning for Skin Cancer Classification: A Study of Model Accuracy, Generalization, and Ensemble Learning

Skin cancer is one of the most common types of cancer, primarily caused by unmutated DNA changes influenced by both environmental and genetic factors. Early and accurate identification is crucial for reducing mortality and improving treatment outcomes. This study utilizes the HAM10000: MNIST dataset, which consists of 10,015 high-resolution dermoscopic images, to evaluate various CNN models. The images were preprocessed and standardized to a consistent resolution to ensure uniformity. Among the models tested, Inception-ResNet-V2 exhibited the lowest accuracy at 51.22%, while VGG19 achieved the highest accuracy at 94.26%. This was followed by DenseNet121 at 93%, Xception at 93%, and ResNet50 at 92%. To further enhance predictive performance, an ensemble learning technique was employed, combining VGG19, DenseNet121, Inception-ResNet-V2, ResNet50, and Xception, resulting in an impressive accuracy rate of 98%. These findings highlight the potential of deep learning and ensemble methods to significantly improve early skin cancer detection, paving the way for more reliable and effective clinical decision-making.