Research On Dermoscopic Images Segmentation Algorithm Based On Deep Learning

Skin disease is a common disease in our life,which can easily affect people’s health.If skin disease is not diagnosed and treated in time,it will have a great impact on the patient.Therefore,accurate diagnosis of skin diseases is very important.However,previous dermoscopic images were diagnosed by professional dermatologists and the lesion area was manually segmented,which is time-consuming and the diagnosis results are subjective.Therefore,computer-aided diagnosis with rapidity and objectivity is particularly important,it can assist doctors to accurately segment dermoscopic images and improve diagnostic efficiency.Nowadays,medical image segmentation methods based on deep learning are widely used in the segmentation task of lesion area.However,the existing algorithms have problems such as low accuracy of the segmented image,blurring,and gradient explosion or gradient disappearance as the depth of the network deepens.Therefore,this paper aims to research an effective segmentation method that can assist doctors in diagnosing skin diseases.The main work content of this paper is as follows:1、Analyzed the segmentation process of dermoscopic medical images based on deep learning,a dermoscopic medical image segmentation experiment based on FCN-8s network,U-Net network,Seg Net network and Res UNet++ network was designed.In the FCN-8s network and Seg Net network experiments,the migration learning method was used,and the VGG pre-trained model was used as the encoder.In the UNet network experiment,the Same convolution is used,which is different from the original network,and transposed convolution is used for upsampling.Through experimental analysis and comparison of the segmentation effect and shortcomings of the four networks.2、Through the comparative analysis of the existing algorithm experiments,drawing on the network architecture of Res UNet++,a dermoscopic medical image segmentation algorithm based on the improved U-Net network was proposed,and improvements were made in four aspects:(1)The residual connection is introduced in the encoding stage,and the convolutional block is replaced by a residual block,effectively alleviate the problem of gradient disappearance.(2)Also in the encoding stage,a lightweight CBAM attention module that can pay attention to both channels and spaces is introduced,improve the network’s attention to the lesion area.(3)A dense hollow space pyramid pooling module is added to the connection of the codec and the final output layer to obtain rich multiscale information,and the number and expansion rate of the expansion convolution in the module are modified.(4)A skip connection is designed to fuse the outputs of different levels of decoders,realize the fusion of different semantics of feature maps of different sizes.3、This article uses the data set HAM10000 from the international skin image official website ISIC,a comparison experiment and ablation experiment were carried out on the improved network.The experimental results show that the network model has been improved compared with other networks in terms of accuracy rate,deiss similarity coefficient,average intersection ratio and precision rate coefficient,and has achieved competitive results.