Performance Evaluation of ResNet50 and MobileNetV2 in Skin Cancer Image Classification with Various Optimizers

Skin cancer is an abnormal growth of skin tissue that can affect some or all layers of the skin. Melanoma is a dangerous form of skin cancer. Early prevention of skin cancer is especially important for high-risk groups. Typically, dermatologists use a biopsy for diagnosis, which involves taking a sample of skin tissue for laboratory examination. However, biopsies are expensive and can cause scarring. Therefore, a computer-based diagnostic alternative is needed to help diagnose skin cancer more quickly and accurately through image analysis. This study evaluates the performance of two deep neural network models, ResNet50 and MobileNet, in detecting skin lesions using three different types of optimizers: SGD, RMSProp, and Adam. This research aims to analyze how the choice of optimizer affects the prediction accuracy and loss value of the two models. Experimental results show that the use of the RMSProp and Adam optimizers generally provides better performance compared to SGD, especially in increasing accuracy and reducing loss values. On the ResNet50 model, the Adam optimizer achieved the highest accuracy of 0.88 with the lowest loss value of 0.4674, while on MobileNet, RMSProp gave the best results with an accuracy of 0.88 and a loss value of 0.3884. Confusion matrix analysis shows that the model has quite good abilities in detecting malignant and benign cases. This research concludes that choosing the right optimizer greatly influences model performance, and MobileNetV2 shows better generalization ability than ResNet50 in this experiment.