Fine-tuned Transfer Learning Models for Multi-class Classification of Melanoma Skin Cancer

Melanoma skin cancer is the deadliest disease that needs detection and treatment at early stages. Systems have been proposed in the literature for melanoma skin cancer detection with its types through standard Machine learning & Deep learning mechanisms. These models still need enhancement to improve the detection rate and reduce the false positives to avoid future losses or problems related to the skin. Transfer learning helps use the existing pre-trained weights on a specific task for other use cases. The transfer learning layers can be retrained by keeping some with the same weights and changing others. This study proposes a new approach to improve melanoma class identification, leveraging the strengths of transfer learning models to improve diagnostic precision. Our methodology involves optimizing four proposed CNN architectures - VGG-16, ResNet-50, Inception V3 & VGG-19 to extract relevant features from dermatological images. We maximize feature extraction and classification outcomes by strategically selecting and fine-tuning convolutional layers. Our evaluation of a comprehensive dataset of melanoma images reveals improved results, surpassing existing methods by a notable margin. The primary aim is to compare different network performances of the pre-trained and fine-tuned networks to enable the selection of the best model for melanoma classification. Standard performance metrics are used to compare the results of the fine-tuned architectures proposed in the paper. ResNet-50 yields the highest accuracy of 98% for multi-class classification, and other fine-tuned architectures have also shown satisfactory performance.