Skin Cancer Classification Based On Convolutional Neural Network

Abstract

Skin cancer is a kind of cancer that is usually diagnosed by images from dermoscopy. In recent years, researchers have attempted to utilize deep learning technology, especially convolution neural networks (CNN), in the recognition of skin cancer images. Many CNN models have already performed great applicability, like DenseNet, Inception, and Xception. This paper carried a comparative experiment on ISIC 2019 challenge dataset which includes 25,331 skin cancer images of 8 different kinds. On the classification task on the ISIC 2019, it introduced 6 models, VGGNet19, ResNet50, ResNet152, DenseNet201, Inception-v3, and Xception, then conducted a comparative analysis of their performance involving 2 methods (data enhance and transfer learning) and 2 optimizers (Adam and SGD), aiming to explore the impact of different methods and structures on the accuracy, in order to find traits for potential models of higher accuracy. In the 24 groups of results, Xception with data enhance, transfer learning (pretraining) and Adam optimizer had the highest accuracy of 83.8%, while VGGNet19 without transfer learning had the lowest of 66.67%. The influence of transfer leaning is positive on all models, both on accuracy and training time; similar to Adam optimizer, except for a noticeable enhancement effect on Inception-v3 and Xception. The data enhance method applied in this paper had a weak, non-directed impact. Possible reasons for this phenomenon are discussed in depth in the study.