Computer-aided Diagnosis System Of Skin Diseases Based On Deep Learning

With the popularity of computers,computer-aided systems have been applied to various fields,bringing great convenience to people.Thanks to the abundance of com-puting resources and the accumulation of various data,deep learning demonstrates supe-rior performance and has achieved outstanding results in many areas.Computer-aided diagnosis system is a hotspot in recent years.In this paper,image-based dermatology auxiliary diagnosis system is mainly studied.In this paper,the diagnosis system based on random forest algorithm is studied.First extract image features,and then make classification.The selection of feature com-bination and the number of decision tree in random forest is explored.In order to avoid the trouble of selecting features and improve the performance of the model,this pa-per studies the diagnosis system based on ResNet and shows that the pre-trained model greatly improves the performance of the networks.In order to further improve the performance of the deep neural networks,in this pa-per we propose a network integration framework named Channel-ResNet.It can achieve better results than simple integration.Appropriate image pre-processing methods are applied and networks are integrated through multiple channels.Experiments are carried out on two kinds of data:dermoscopic images and skin surface photos.For dermoscopic images,there are three diseases involved:melanoma,nevus and seborrheic keratosis.The metric is the mean value of the AUC for the melanoma and seborrheic keratosis classifications.In this problem,the random for-est can achieve a result of 0.7333 on the test set,while the ResNet can reach 0.8714.And the result of Channel-ResNet is 0.9168.For skin surface photos,there are four diseases in volved:infantile eczema,infantile heatrash,baby subitum and baby vari-cella.The metric is accuracy of the classification.In this problem,the random forest can achieve a result of 63.2%on the test set,while the ResNet can reach 79.2%.And the result of Channel-ResNet is 83.2%.Our proposed network integration framework has significantly improved the results.