Classification And Recognition Of Melanoma Dermatosis Based On CNN

In recent years,the rapid development of computer vision technology has promoted the invention of new technologies in medical field.In this paper,we use computer vision technology to solve the classification problem of malignant melanoma images under skin mirror.According to the data,melanoma accounts for 5% of skin diseases.It is also the highest death rate in skin diseases,accounting for 75% of the total mortality of skin diseases.If the early diagnosis of melanoma lesions and through treatment,the survival rate is up to 97% within 5 years,which makes the early skin mirror image of melanoma diagnosis significant,help to avoid thousands of patients miss the treatment due to delay.Deep learning has outstanding performance in computer vision,especially in the field of image classification and recognition.But the process of deep learning requires a large number of data to do training sets,and the data set in this paper is limited,and there is a data imbalance problem,which increases the difficulty of the task of the skin mirror image classification.Aiming at the above problems,this paper uses the skin mirror image as the research object,uses the deep convolution neural network(CNN)algorithm,simultaneously through the data enhancement and the improved loss function method to realize the classification and recognition of the skin disease image.Finally,integrated learning is used to further improve the generalization of the convolution neural network.In this paper,we mainly study the detection and classification of malignant melanoma skin diseases under dermatoscope.The research contents are mainly divided into four parts.First,based on the features of dermatoscopy,the deep convolution neural network is used to classify and identify the image of benign dermatosis and the image of malignant melanoma.Second,because of the small amount of data and the problem of data imbalance,this paper,based on the deep convolution neural network,aiming at the problem of data imbalance,first uses a special data enhancement method,and then combines the improved loss function: Focalloss,to further improve the classification accuracy of the deep convolution neural network.In a limited and unbalanced training set,we improve the generalization of the final training model.Through the comparison of experimental results,it can be seen that the effect of neural network classification is obviously better than that of the ordinary neural network classification with the enhancement of data and the improvement of the loss function.Third,after the improved deep convolution neural network,the training results of multiple neural networks are integrated through the Boosting method in integrated learning.It can be seen from the contrast experiment that the effect of the multi neural network model combined with integrated learning is obviously better than the effect of the single neural network,and the network generalization is further proposed.High.Fourth,this paper joins the two integrated learning,and integrates the Boosting algorithm integrated deep convolution neural network with other trained neural network models to carry out similar bagging algorithm integration.The experimental results show that the effect is better than the single Boosting algorithm integrated neural network,and it is proved that the integrated learning can be greatly improved.The generalization of deep neural network makes the final training model get better results.The results of this study are remarkable.The accuracy,AUC,AP,sensitivity,specificity and other classification indexes are all improved in the first place of the competition in 2016,at least0.5 points can be raised,and up to 6.6 points can be raised at most.