Study Of Segmentation Method For Skin Lesion Images Based On Adversative Transfer Learning

Skin disease is one of the most common diseases in human life,and pigmented skin disease like melanoma is a skin cancer with a high mortality rate.The use of dermatoscopy imaging technology can clearly show the characteristics of lesions to assist doctors in treatment.In this process,whether professional doctors perform cutting surgery or laser resection with precision instruments,it is necessary to achieve accurate segmentation of the edge of skin lesions.With the development of computer vision technology,the computeraided diagnosis system can learn from the experience of professional doctors to diagnose and analyze dermatoscopic images,and improve the efficiency and accuracy of skin lesion area segmentation in dermatoscopic images.In recent years,deep convolutional neural network has achieved good results in the field of medical image segmentation.According to the form of segmentation network,it can be divided into FCN,U-Net and GAN-based networks,gradually expanding from supervised learning to unsupervised learning,with faster and faster segmentation speed and higher and higher segmentation accuracy.However,in the process of skin lesion image segmentation,these segmentation methods are still affected by such factors as fewer professional labeled samples,large diversity of lesion range,low contrast between lesion area and background,and binary feature images dependent on pixel level.In order to effectively solve the segmentation problem of skin lesion images,this paper proposed a segmentation algorithm based on improved Mask R-CNN based on the idea of adversity-transfer learning and common segmentation methods of skin lesion images.Based on this,an adversarial transfer learning segmentation model(Adversarial Transfer in Mask and Region-based Convolutional Neural Network,AT-MRCNN)is proposed to solve the problem of few training samples and incomplete edge features segmentation.To effectively achieve the objectives in the proposed model,the main contents of this article are as follows:Firstly,the development of convolutional neural networks and three segmentation networks commonly used in the field of medical image segmentation are analyzed.The first one is based on full convolution,the second one is based on coding-decoding structure,and the third one is based on generative adversarial structure.At present,more and more researches in the field of medical image processing use the third network adversarial learning method,and this paper proposes a segmentation model based on adversarial.Secondly,Mask R-CNN was used as the basic network to improve,and Res Next-101 with better performance was used as the backbone network,so that the number of computing parameters would not be greatly increased while the number of network layers was deepened.Expanding convolution is added at the end of the output layer of feature extraction to fuse the feature information of the context and enlarge the global feature receptive field.Mask R-CNN is a network with both detection and segmentation.According to this feature,the weight parameters of the last layer of detection branch are mapped to the last layer of segmentation branch by assigning transfer function,and the detected skin damage position is learned before segmentation output to improve segmentation accuracy.The improved segmentation network named MRCNN was added to the generator part of the adversation model,and a discriminator of convolutional neural network structure was designed,and softmax layer was used as the classification output at the end of the discriminator.In this study,an adversarial transfer learning model AT-MRCNN is proposed based on the segmented network MRCNN and the discriminator.Finally,in order to better verify the segmentation effect of AT-MRCNN model,an open data set was used for experimental analysis,and the experimental data were all from the Skin Image Association ISIC.The experimental data set is divided into source domain and target domain.The source domain data set contains a small number of images labeled by professional physicians,and the target domain data set contains a large number of unlabeled dermatoscopic images.The experiment first verifies the performance of the improved MRCNN segmentation network,and then verifies the performance of AT-MRCNN segmentation model.After comparative analysis of several commonly used segmentation methods,the AT-MRCNN model proposed in this study can effectively complete the accurate segmentation of skin lesions,and the similarity coefficient between the predicted segmentation results of the model and the real labeling reaches 87.7%,which verifies the reliability and effectiveness of the segmentation method proposed in this study in the field of skin lesions segmentation.