Research On Skin Disease Image Segmentation And Classification Based On Deep Learning

Skin cancer is a kind of cancer that poses a great threat to people’s life safety.Melanoma,the most deadly skin cancer,is a pigmented malignant lesion with characteristics of difficult detection and high mortality.Early detection and early treatment of it is more important.In nowadays in the process of diagnosis and treatment,however,usually rely on the doctor’s clinical experience in the professional,by looking at the size of the lesion area,such as color,texture information to judge types of lesions or lesions area,this relatively subjective judgment in the face of the diversity of the skin disease prone to misdiagnosis phenomenon,for will cause the waste of medical resources.Therefore,it is very important to help doctors accurately determine the type and area of lesions through computer-assisted dermatology diagnosis and treatment,which can also make serious skin diseases such as melanoma receive faster and more professional treatment and improve the survival rate of patients.Based on the above background,this paper studies the segmentation and classification of skin disease images based on deep learning.The main contents are as follows:(1)Aiming at the phenomenon of blurred boundary of skin disease image and different size of lesion area,a skin disease image segmentation algorithm was proposed based on U-NET convolutional neural network.Res Net50 network was used to replace the coding structure in the model to obtain the feature layer with multi-level semantic information.Then upsampling at different magnification rates was used to unify their sizes and feature fusion was carried out through channel dimension connection.The output of the previous operation is passed through the information fusion module to improve the information loss caused by the direct convolution operation.Finally,more important features are screened through the collaborative attention module.The experimental results showed that in the official data set of ISBI2016 Melanoma Lesion Detection Challenge,the Jaccard Index segmentation performance Index reached87.17%,2.37% higher than the U-NET model,and other indicators also achieved comprehensive surpass.(2)Aiming at the poor segmentation effect of skin lesions with hair occlusion,a skin lesions image segmentation network based on dual attention mechanism was designed.Firstly,feature layers of different sizes are obtained for the image through Res Net50 network,and the extracted features are fused with multiple features.Then,multiple features are extracted after entering the next encoding and decoding path.Finally,the final output is obtained through spatial and channel attention module.Comparative experiments and ablation experiments were conducted on ISBI2016 data set,and the experimental results showed that the images covered by hair and other objects had excellent segmentation results.In addition,the experimental indexes were 96.19% accuracy,93.32% sensitivity,97.32% specificity,93.26% Dice coefficient and 87.36% Jaccard index,which were better than the existing model and the model algorithm proposed in the competition.(3)A lightweight skin disease image classification network was designed for the similarity between different types of skin diseases,and the types of lesions were prone to misjudgment.The depth separable convolution is used to reduce model parameters and the network depth is deepened by stacking bottleneck blocks with residual structure.In addition,the feature layer with richer feature information is obtained by redesigning the convolution layer and pooling layer in the down-sampling process.The proposed model achieves 86.03% classification accuracy on HAM10000 data set,which is superior to existing models in classification accuracy and has certain advantages in the number of model parameters.