Research On Skin Lesion Segmentation Method Based On Convolution Neural Networks

The incidence of skin lesions worldwide is increasing year by year,posing a huge threat to human life and health.If skin lesions such as melanoma can be diagnosed and treated as soon as possible,the survival rate of patients can be greatly improved.However,manual diagnosis by clinicians alone cannot meet the public’s requirements for timeliness and convenience of diagnosis.The application of computer-aided diagnosis methods to accurately segment the boundaries of skin lesions is of great significance for assisting doctors in diagnosis and treatment.Most of the existing skin lesion segmentation models have insufficient feature information extraction due to the small acquisition field of the model and the lack of global context,which makes it impossible to accurately segment the foreground and background of skin lesions.In response to the above problems,this paper proposes two models based on the attention convolutional neural network.The main research content of the paper can be summarized as follows:1.An end-to-end multi-scale attention convolutional network model CSARM-CNN(Channel & Spatial Attention Residual Module)is proposed to accurately segment the lesion area in dermoscopic images.Among them,a novel attention learning module CSARM is designed,which embeds the convolution module and the attention module at the same time to further improve the feature representation of the model.The CSARM-CNN model uses U-net as the basic structure,uses CSARM blocks and convolution operations to extract features from images.In the encoder path,build image pyramids to feed multi-scale inputs.In the decoder path,a local prediction map corresponding to the multi-scale input image is generated,and the multi-output cross-entropy loss function is used to promote the training of the model.2.An end-to-end deep convolution generation confrontation network model ACG-Net(Attention Convolution Generative-Adversarial Network)is designed to further realize the accurate segmentation of skin lesion images.Combining the idea of generating a confrontation network,the model is composed of two network structures:generator(G)and discriminator(D).G and D are optimized through confrontation training.The generator that uses CSARM-CNN as the model uses skin cancer images as input,and the discriminator is a network model with alternating convolution and global convolution.The skin lesion image and the label image and the generated image are respectively stitched together as input data.Finally,the output is the probability of the segmentation label map of the skin lesion image,thereby optimizing the parameters in the entire network structure.3.On the ISIC-2017 dataset,a large number of ablation experiments and statistical hypothesis tests have verified that the CSARM block has excellent feature information extraction capabilities in the skin lesion segmentation task.And comparing the performance of CSARM-CNN with other latest skin lesion image segmentation algorithms,experiments show that the accuracy and specificity are significantly improved.In order to verify the robustness of the model,the model is tested on another publicly available data set PH2,and competitive results are also obtained.Finally,in order to verify that ACG-Net combined with a generative adversarial network can effectively improve the segmentation performance of the model,a comparative experiment is carried out between CSARM-CNN and ACG-Net.The model is evaluated by six evaluation indicators including accuracy and Dice coefficient.The results show that ACG-Net’s other five indicators except for specificity are all higher than CSARM-CNN,achieving competitive experimental results.