Research On Skin Lesion Recognition Algorithm Based On Deep Learning

Objective: The late fatality rate of malignant skin lesions represented by melanoma is extremely high,which seriously threatens people’s health and life.If the skin lesion can be detected and treated in time,the survival rate of the patient can be greatly improved.At present,dermatologists usually take the corresponding dermoscopic images for analysis and then diagnose whether a patient has skin lesion.However,this method is time-consuming and the doctor’s work is very large.In addition,due to the subjectivity of the diagnostic process,misdiagnosis and missed diagnosis may occur.With the development of deep learning technology,computer-aided diagnosis,including skin lesion segmentation,feature extraction and classification,has gradually received attention.However,due to the skin lesion image characteristics,it faces many challenges in practical research and application.Based on this,this study focuses on the automatic and accurate segmentation of skin lesions to improve the accuracy of skin lesion segmentation.At the same time,the application of skin lesion segmentation in the classification task is discussed,which provides research ideas for the combination of segmentation work and classification task.Methods and results:(1)Aiming at the problems of low contrast,blurred edges,noise interference and other problems in the skin lesion segmentation,this study proposes an encoder-decoder network structure based on residual connection and channel attention mechanism,Res SE-UNet.By introducing residual connections in the sub-modules of U-Net encoder and decoder,the network transmits shallow information to the deeper layers of the network and strengthens the network feature transmission.In the skip connection part,the channel attention mechanism is used to strengthen the network feature refinement ability and reduce the interference of irrelevant information.The group normalization method is used instead of the commonly used batch normalization method to solve the problem of increasing the error caused by batch normalization when the batch size is too small.The segmentation performance of the model was verified on the ISIC 2018 dataset,and the Dice coefficient,Jaccard coefficient and accuracy were 90.23%,83.86% and 95.98%,respectively,which are improved by2.19%,2.93% and 1.21% compared with the basic U-Net.(2)Based on some shortcomings of Res SE-UNet in skin lesion segmentation,this study continues to optimize on the basis of Res SE-UNet,and proposes a convolutional neural network model based on multi-scale feature extraction and channel spatial attention mechanism.The model removes the residual connection introduced in Res SE-UNet,and then adds two atrous spatial pyramid module composed of dilated convolution with different expansion rates in the bottleneck part of U-Net for multi-scale feature extraction,so as to improve the network segmentation effect while controlling the network computation amount.In the skip connection part,the dual attention mechanism is used to realize adaptive feature refinement from two dimensions: channel and space.The group normalization is still used to improve the segmentation performance of the network in minibatch processing.The method was evaluated using the ISIC 2018 dataset and discussed the dilated rate setting in the atrous spatial pyramid module,whether to remove the global average pooling layer,and the number of channels per group during group normalization.Finally,the Dice coefficient,Jaccard coefficient and accuracy obtained by the model were90.39%,84.02% and 95.98%,respectively,which are 2.35%,3.09% and 1.21%higher than the basic U-Net.In addition,compared with Res SE-UNet,the network further improves the segmentation effect of skin lesion areas.(3)In order to prove that feeding the mask image obtained by the segmentation work into the classification network together with the skin lesion image can achieve a better classification effect compared with skin lesions image directly,this study uses the trained segmentation model to obtain the mask of each image of the ISIC 2017 dataset.And then,the original image,the image after the original image is added with the segmentation mask,and the image after the original image and the segmentation mask is concatenated are sent to the classification network model Res Net34 for skin lesion classification.Finally,the results of each evaluation index are obtained.From the results,the average diagnostic accuracy obtained by adding mask as network input was the highest,followed by the concatenated operation,and the classification accuracy of the original image directly input was the lowest.Conclusion: It can be seen from the results that the network model can strengthen the learning of feature information by the network through the residual module,and improve the processing ability of blurred lesion edges and low-contrast lesion areas.The use of attention mechanism can make the network focus on the lesion area instead of the non-lesion area,reduce the interference of noise such as hair and black artifact,and improve the segmentation accuracy.The group normalization method reduces the error caused by network training and further improves the segmentation performance of the network model.According to the experimental results,it can be concluded that the atrous spatial pyramid module can expand the receptive field of the model and strengthen the refinement of the lesion edge by parallel operation of dilated convolution with different dilation rates,while controlling the amount of calculation.The dual attention mechanism can further improve the model’s filtering of noise,such as hair and artifacts.It can be concluded that the integration of segmentation mask results into the input image of the classification network makes the classification network pay more attention to the lesion area rather than the background area,thereby improving the diagnostic ability of the classification network for different lesions.This also proves the important value and research significance of skin lesion segmentation in computer-aided diagnosis.