Skin cancer diagnosis using hybrid deep pre-trained convolutional neural networks

Abstract

As a variant of skin cancer, melanoma represents a substantial menace to the health and overall well-being of individuals. Statistics reveal that 55% of skin cancer patients succumb to this particular disease. However, early detection plays a crucial role in reducing mortality rates and saving lives. In the past several decades, there has been a rise in the adoption of deep learning algorithms, capturing the interest of researchers working in this field. One popular method involves utilizing pre-trained deep neural networks. In this study, a hybrid approach is employed to extract features from melanoma images. This approach integrates the utilization of pre-trained architectures, including AlexNet, ResNet-50, and GoogleNet. During the transfer training phase, these networks are fine-tuned to detect skin cancer by adjusting the learning rate. Subsequently, the maximum relevance minimum redundancy (MRMR) algorithm is employed to select optimal features based on the concepts of minimum redundancy and maximum relevance in order to minimize feature redundancy and enhance classification accuracy. The bagging technique is employed for the classification of various skin cancer types. The experimental results demonstrate the success of the suggested approach, yielding 98.9% accuracy. Furthermore, the results indicate the superiority of this method according to precision, recall, and F1-score in comparison with existing algorithms.