Enhanced Classification of Skin Lesions Using Fine-Tuned MobileNet and DenseNet121 Models with Ensemble Learning

This study presents a deep learning approach for early detection of melanoma, one of the most dangerous skin cancers. In this article, all pre-trained models of the Keras library are trained with the ISIC skin cancer dataset available on Kaggle and the accuracy of each model is analyzed in detail. With the results obtained from the trained models, the models were fine-tuned to further optimize the performance of each model. After re-evaluation with fine-tuning, the accuracy rates were compared: DenseNet121 and MobileNet were found to be the two best models with high accuracy among the fine-tuned models. As such, these two models were combined in an ensemble approach to achieve a better overall accuracy. The skin cancer detection rate obtained with this ensemble approach is 93.03%. Therefore, the deep learning-based ensemble method appears to be a reliable and powerful technique that can be used to diagnose serious diseases such as skin cancer. This model can be used to provide a powerful support system with great potential to assist dermatologists in the early detection phase by easing workload and improving patient outcomes.