Research On Skin Disease Detection Method Based On Convolutional Neural Network

Cancer is already one of the biggest culprits that threaten human life,and skin cancer is the most overlooked cancer.If skin cancer can be diagnosed and treated early,survival rates can be greatly increased.Conventionally,the detection of skin diseases is first performed through dermoscopy to generate high-definition dermoscopy images,and then identified by professional dermatologists.However,this lesion detection process relies heavily on the experience of experts and is highly subjective.Therefore,it is very necessary to develop a more accurate and reliable system to replace manual diagnosis.However,due to the four reasons of limited data set,imbalanced data set,difficulty in feature extraction,and small skin lesion area,traditional classification models often fail to achieve good results in detecting skin diseases,a large amount of work has been done to address the first three issues,while relatively little work has been done to address the fourth issue.Therefore,this thesis mainly explores the method to solve the problem of small skin lesion area,and a related model is established to improve the performance of lesion detection.The main contributions of this thesis are as follows:(1)First,the existing skin disease detection methods are summarized,and the existing skin disease detection methods are mainly divided into four categories according to the existing problems.For the problem of limited data sets,methods such as pre-trained models are mainly used;for the problem of data set imbalance,methods such as sampling,data synthesis,and cost sensitivity are mainly used;for the problem of difficult feature extraction,methods such as ensemble learning,attention mechanism,meta-information,and multi-task learning are mainly used;for the problem of small skin lesions,mainly by introducing segmentation information and other methods.(2)There are a lot of solutions for the first three problems,but most of the existing work ignores the small area of skin lesions in dermoscopic images.Aiming at this problem,this thesis designs a two-stage detection method.In the first stage,an improved fully convolutional residual network(IFCRN)is used to segment the skin lesion image.Then,in the second stage,the area containing the skin lesions is cropped according to the segmentation masks,and enlarged to a uniform size,finally,a deep residual network(DRN)classification network is used to classify and detect the enlarged skin disease image.Experiments show that training the classification network based on the segmentation results instead of the original skin disease images can effectively prevent the classification network from being interfered by other structures and artifacts in the image during the training process,and also enlarge the skin lesion area to generate more distinguishing features,obtains a better recognition effect.(3)In order to solve the problem that the segmentation network cannot be trained well due to the difficulty of obtaining segmentation labels,a double-branch attention convolutional neural network(DACNN)based dermatological detection framework is proposed.First,the DACNN framework explicitly includes upper and lower branch networks adopting the same structure.Each branch network adopts the ARL(Attention Residual Learning)structure to enhance the ability to extract the features of the lesion area,and uses the LLN(Lesion Location Network)to connect the upper and lower branch network structure.Introduceing an attention mechanism,the LLN locate lesion area,crop and enlarge the lesion area as the input of the lower branch network.The lower branch network can focus on the lesion area,so that the lesion area can still be targeted without the need for lesion segmentation labels.Experiments show that compared with several existing typical skin lesion detection methods,the performance of DACNN has been greatly improved.