Al-Saedi, Doaa Khalid AbdulridhaSavaş, Serkan2025-01-212025-01-2120242520-8438https://doi.org/10.1007/s42044-024-00210-yhttps://hdl.handle.net/20.500.12587/23569Skin cancer represents a significant global health threat with potentially fatal consequences if left undiagnosed. Early detection is crucial for successful patient treatment, yet accurate identification of skin lesions poses a challenge even for experienced dermatologists. In this context, the development of computer-aided skin lesion classification systems emerges as a promising path to empower dermatologists with the potential for earlier diagnoses and more effective treatment interventions. This study proposes a two-stage approach for early detection of skin cancer. Firstly, eight pre-trained deep architectures were tested on the ISIC dataset using transfer learning and fine-tuning. Secondly, three successful models with the highest accuracy were chosen, and ensemble learning was employed to obtain a final result. The ensemble learning method outperformed individual models, achieving a remarkable ROC AUC rate of 99.96%. DenseNet121 exhibited the highest performance among the individual models, with accuracy rates of 99.75%, 98.2%, and 99.6% for the train, validation, and test datasets, respectively. The promising results hold significant potential for early detection and treatment of skin cancer, a prevalent global disease. These findings could prove invaluable for clinics, offering valuable support to their decision-making processes and enhancing their ability to combat this widespread health concern. © The Author(s), under exclusive licence to Springer Nature Switzerland AG 2024.eninfo:eu-repo/semantics/closedAccessDeep learning; Ensemble learning; ISIC; Skin cancer; Skin lesion; Transfer learningArticle7478580010.1007/s42044-024-00210-y2-s2.0-85207669096N/A