Skin cancer, particularly melanoma, has remained a severe issue for many years due to its increasing incidences. The rising mortality rate associated with melanoma demands immediate attention at early stages to facilitate timely diagnosis and effective treatment. Due to the similar visual appearance of malignant tumors and normal cells, the detection and classification of melanoma are considered to be one of the most challenging tasks. Detecting melanoma accurately and promptly is essential to diagnosis and treatment, which can contribute significantly to patient survival. A new dataset, Nailmelonma, is presented in this study in order to train and evaluate various deep learning models applying transfer learning for an indigenous nail melanoma localization dataset. Using the dermoscopic image datasets, seven CNN-based DL architectures (viz., VGG19, ResNet101, ResNet152V2, Xception, InceptionV3, MobileNet, and MobileNetv2) have been trained and tested for the classification of skin lesions for melanoma detection. The trained models have been validated, and key performance parameters (i.e., accuracy, recall, specificity, precision, and F1-score) are systematically evaluated to test the performance of each transfer learning model. The results indicated that the proposed workflow could realize and achieve more than 95% accuracy. In addition, we show how the quantization of such models can enable them for memory-constrained mobile/edge devices. To facilitate an accurate, timely, and faster diagnosis of nail melanoma and to evaluate the early detection of other types of skin cancer, the proposed workflow can be readily applied and robust to the early detection of nail melanoma.
