Skin Lesion Segmentation And Classification Algorithm Based On SU-SWA And Deep Ensemble Model

Skin cancer is the world‘s fastest growing disease,including melanoma as the representative of the skin cancer is one of the world‘s deadliest cancers,skin cancer in timely diagnosis and treatment,can improve the five-year survival rate of such diseases.Now,dermatologists mainly use dermorscopy technology to diagnose dermatologists to micro-examine the lesions,then observe and analyze the lesions,and finally draw a conclusion.As aging increases,so does the number of skin disease patients.The demand of computer-assisted diagnosis and treatment system of dermatology is increasing day by day,and the auxiliary analysis can relieve the consultation pressure of dermatologists and improve the efficiency of treatment.Since2012,deep learning has been developing continuously,and models based on convolutional neural networks have achieved good results in medical image analysis.In dermoscopy image analysis,there are two main tasks of skin lesion area segmentation and diagnosis.The purpose of the division of the skin lesion area is to assist the doctor to remove the lesions accurately.The purpose of automatic skin lesion area disease diagnosis is to realize the diagnosis based on the dermoscopy image,to achieve the graded shunt of the skin patient,greatly reduce the doctor‘s repeated work and improve the efficiency of the doctor’s diagnosis.Based on the further study of deep ensemble learning and model fusion,in this paper,we improve the application of the two tasks of deep learning model in the case of deomoscopy image with small sample.The main research content is as follows:1)when it comes to the segmentation task,a Separable-UNet model is proposed,which combines the advantages of separable convolution operation with UNet structure.It can fully obtain the correlation between the feature channels and the advanced semantic feature information,so as to enhance the precision of segmentation.At the same time,in response to the problem of overfitting the deep learning model,in this paper,we propose the use of stochastics weighted average strategy(SWA),Weighted average of multiple local optimal points to make the model get better optimization and generalization performance,which improves the segmentation model’s performance.Finally,compared with other skin-loss segmentation frame,this paper from the skin-loss division of the overall frame of image pre-processing,image size selection,segmentation model parameters,after the image processing four parts to be streamlined,so that the skin loss of the frame single picture split time of only 0.044 s.Under the same hardware configuration,it is faster than other skin-damaged segmentation methods.2)In view of the skin lesion area diagnostic task,a global-part convolution neural network model(Global-Part Convolutional Neural Network,GP-CNN)is proposed.The global model is trained by the subsampled dermal mirror image to obtain the global information of the image and the corresponding class activation map(CAM),and the local model is trained by the cam-guided extracted tiles to obtain fine-grained information about the skin mirror image.Finally,by weighting the fusion of global context information with local fine-grained information,the accuracy of skin damage area disease diagnosis is improved.At the same time,a data-transformed ensemble learning(DTEL)is proposed.The accuracy of diagnosis is further improved by the fusion of GP-CNNs different decision information trained by the original image,color constant transformation image and feature saliency transformation image.3)A set of skin image analysis software has been developed for the PC and mobile terminals,respectively,using the programming language of Python and deep learning framework of Keras and Tensorflow.This software integrates the model of skin lesion segmentation and diagnosis,which can realize the real-time detection of these two tasks.