Improving Skin Cancer Diagnosis: A Comprehensive Analysis of Transfer Learning and Ensemble Models

Skin cancer has become a major global health issue, necessitating improved diagnostic methods for better clinical outcomes. This study thoroughly evaluates deep learning methods for classifying skin cancer using dermoscopic images. We compare the performance of popular pre-trained models, including VGG16, ResNet50, Xception, and MobileNet, against a natively trained SqueezeNet model. The results show that VGG16 and ResNet50 models have achieved state-of-the-art accuracy of 94%, with SqueezeNet following closely at 93%. This validates the effectiveness of transfer learning. We suggest using ensemble models that combine pre-trained networks (VGG16, MobileNet, ResNet, and Xception) to achieve a maximum accuracy of 97% using model averaging. Beyond accuracy, our ensemble models excel across Precision, Recall, and F1-score metrics, emphasizing their potential for real-world usage. Additionally, we present ROC curve analysis and confusion matrices to provide clinical insights. This study highlights the capabilities of transfer learning for skin cancer classification while introducing novel ensemble techniques that can profoundly impact computer-aided diagnosis through improved accuracy and robustness.