Melanoma Cancer Dermoscopy Image Classification Based on ResNet Model and Its Optimization

Nowadays, the most prevalent way to detect Melanoma Cancer is dermoscopy examination, however it is a challenging process for doctors which costs a lot of time and hard to precisely analyze. This paper introduces a new method to help doctors diagnose Melanoma Cancer more quickly and with higher accuracy by using a new convolutional neural network (CNN) called ResNet. This image classification model is constructed on thousands of dermoscope images from Kaggle dataset [1], which respective belong to benign and malignant. The NLM (Non-Local Means) filter is applied to filter out distracting information and avoid over-fitting. Six different optimizers models are also utilized, including SGD, RMSprop, Adam, Adamax, Adagrad. In advance, their classification qualities are compared from aspect of accuracy, loss, convergence rate, precision, loss rate, and confusion matrix. In the end, it can be identified that Adamax is the most suitable optimizer for the image classification of dermoscope images in the ResNet neural network structure.