Skin lesion classification by weighted ensemble deep learning

Skin 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.