Skin Lesion Classification based on Deep Convolutional Neural Network

Abstract

Skin cancer is one of the most common cancers, and its early detection has a huge impact on its outcomes. In this paper, the deep convolutional neural network is improved with the idea of transfer learning to classify 7 types of skin lesions that are from ISIC 2017 HAM10000 dataset. First, the skin lesion images are trained with a 3-layer convolutional neural network after preprocessing. Then for InceptionV3, ResNet50 and DenseNet201, remove the output layer of each original network, add new pooling layer and full connection layers to different networks respectively. After that, combine some of the convolution layers and pooling layers with the new pooling and full connection layers to form three new improved models, based on the original deep conventional networks. Finally, the training parameters which are from ImageNet network are fine-tuned on new improved InceptionV3, ResNet50 and DenseNet201 to finish the classification. The experimental images' size is 224* 224, and the experiments turn out that three improved networks get better results, and the improved ResNet50 gets the best which accuracy is 86.69%. Its accuracy is 3% higher than the comparable other methods.