Research On Skin Image Classification Based On Mutual Deep Learning

Today,with the development of medical conditions,skin cancer is a major public health problem and an important cause of human death.Skin cancer is divided into melanoma and non-melanoma skin cancer.Melanoma is the most invasive skin cancer and the most deadly skin cancer.The 5-year survival rate of advanced melanoma is less than 40%.Non-melanoma skin cancer is the most easily diagnosed malignancy.Therefore,skin cancer is still one of the cancers that we need to focus on.Dermatoscopy is the visual inspection technology.It can enlarge the skin and eliminate surface reflections.Automatic classification of skin lesions in dermoscopy is an important method to improve the diagnostic performance and reduce the death of melanoma.Recently,deep learning has had great success in the classification of dermatological images.Although deep learning has proven to be superior to traditional manual characterization methods,the task of classifying skin lesions remains challenging.One is that there are too few images of skin lesions,but deep learning needs a lot of training data.Secondly,the existing classification algorithms for skin lesion images have certain limitations.The accuracy of skin lesion image classification should be further improved due to these two reasons.This thesis makes the following research on the above problems:(1)Aiming at the problem of too few images of skin lesions,data augmentation is carried out in this thesis.On the one hand,it can prevent the occurrence of overfitting phenomenon,and on the other hand,it can improve the classification accuracy of the network.For the preprocessed images,this thesis proposes a skin lesion image classification method based on the SE-Inception-Resnet-v2 network,and compares it with existing classification methods.Experiments prove that this classification method works better.(2)For the SE-Inception-Resnet-v2 skin image classification method proposed in this thesis,a classification method of skin lesions image based on mutual deep learning model is proposed.The method uses two SE-Inception-Resnet-v2 networks,and each SE-InceptionResnet-v2 network is trained with the mimcry loss based on the Kullback Leibler divergence and a supervised learning loss.The mimcry loss based on the Kullback Leibler divergence is added to the loss of each network,so that the two networks can learn from each other and learn cooperatively,thus improving the classification performance of each network.Experiments show that the model proposed in this thesis can accurately classify the ISIC skin lesions classification data set in 2016.Compared with the existing image classification methods of skin lesions,the model has higher accuracy.