Research On The Prediction Of Skin Cancer Based On Transfer Learning Of Convolutional Neural Network

Malignant skin disease is one of the deadly cancers,and early diagnosis is very important,which can significantly improve the survival rate of patients.In recent years,the Convolutional Neural Network(CNN)method has shown great advantages in the field of medical imaging.However,it is difficult to obtain large-scale medical images with annotations,and training time is long,while small-scale medicine data training from scratch can easily lead to over-fitting problems.Therefore,this thesis proposes a CNN-based secondary migration method to identify skin images.Using the public dataset ISIC2018 as the target domain for the first migration,and the target domain for the second migration is the public data set ISIC2017.The main research contents of this thesis include the following two aspects:The pre-trained CNN model on the ImageNet dataset is first migrated to the skin image.According to the characteristics of the skin dataset ISIC2018,this thesis uses three different convolutional networks,VGG19,Inception v4 and DenseNet169,and reconstructs the corresponding fully connected layer,and then loads the ImageNet pre-trained weights as the initialization weight of the convolutional network,training for ISIC2018 data migration.Due to the small number of data samples,the network training uses the method of updating part of the network layer weights,and the remaining layer weights remain unchanged to extract the basic features of the image.The experimental results show that the accuracy of the three models VGG19,Inception v4 and DenseNet169 after transfer training are 90.25%,88.31% and 91.50% on the ISIC2018 dataset.This thesis selects the best model DenseNet169 as the training network model for secondary migration.Because the ImageNet dataset and the ISIC2018 dataset feature are not so similar,the effect of post-migration training and prediction will be affected to some extent.Therefore,based on the model of DenseNet169’s first migration training,this thesis migrates the weights after ISIC2018 training to a higher similarity feature,and performs the second learning training on the skin dataset ISIC2017,and tests the prediction effect of the DenseNet169.The experimental results show that after two migration studies on the DenseNet169 network model,the prediction accuracy of skin diseases is improved.