Research On Skin Lesion Image Segmentation Method Based On Convolutional Neural Network

Air pollution is becoming more and more serious,and the incidence of skin diseases is also increasing,which poses a great threat to human life.Malignant and fatal skin diseases,led by melanoma,can be diagnosed and treated at an early stage,and the survival rate of patients will be greatly improved.At present,clinical diagnosis only by professional doctors can not meet the requirements of timeliness and convenience of diagnosis.Computer aided diagnosis and treatment can be helpful for doctors to diagnose pigmented skin diseases.In this paper,the U-Net++algorithm is improved and an efficient and accurate skin lesion image segmentation model DAF-Unet ++ is proposed.The specific contents are as follows:Firstly,aiming at the problem of inaccurate segmentation results caused by fuzzy and fragmentary wound edges of some skin lesion images,the extended convolution block using standard convolution and extended convolution alternately was used to replace the standard convolution block of the original network.A variety of features extracted from different receptive fields were fused to enhance the segmentation accuracy of the network.Secondly,the application of intermediate convolution and nonlinear excitation functions in convolutional networks results in the loss of spatial details of image features in high-dimensional output features.The flexible attention mechanism was used to filter and combine spatial information from different dimensions,and output important feature information of focused skin lesions.In the image of skin lesions,the number of pixel samples in the region to be segmented is far less than the number of background pixel samples,resulting in data imbalance.Focal Loss function was introduced to make the network model pay more attention to pixel samples of the lesion area in the training process.Finally,after the original network output convolution layer,a layer of convolution is added to connect the multi-level output feature information to realize multi-scale feature information fusion,so that the improved network model can learn the semantic information features of different levels.In order to prove the performance of the skin lesion segmentation method designed in this paper,experiments were carried out on the open source dataset ISIC Challenge 2017.Experimental results show that the proposed DAF-Unet ++algorithm performs well in the segmentation results,accurately dividing the contour and edge details of the lesion area,and the accuracy and sensitivity of the segmentation results reach 93.4% and 85.5% in ISIC2017 data set,respectively.Compared with many existing excellent skin lesion image segmentation algorithms,it has better segmentation performance and meets the requirements of medical assisted diagnosis.