Classification of Dermoscopy Skin Images with the Application of Deep Learning Techniques

The aid-computer assistant is important for the improvement in the diagnostic of skin lesions, principally to detect melanoma, which is the most dangerous skin cancer type. The principal goal of the aid-computer assistant is to detect melanoma in its initial stage because the probability of the five-year survival ratio is over 95%. With the purpose of diagnosed early, in this paper we propose an algorithm based on deep learning to classified dermoscopy images with two types of skin lesions, which are melanoma or malignant cancer and benign cancer. We use AlexNet architecture modified to attend our problem, which was trained twice times with 2000 and 3000 images, divide into three sets (training, validation, and test set). We analyze three types of optimizers in order to improve the process of learning, where the best result was obtained by SGD with an accuracy of 99.79% for the training set, 81.50% for the validation set, and 79.17% for the test set in the scenery of 4000 images.