Skin Wound Image Segmentation Based On Feature Augment And Interactive Neural Network

Skin is the first line of defense to protect the human body from the outside world.So it is extremely vulnerable to injuries and the invasion of microorganisms,and skin wounds are formed.The use of skin wound images for wound segmentation is a non-invasive method,which can accurately measure the wound area and assist doctors in the diagnosis and treatment of wounds.Therefore,skin wound image segmentation has important clinical significance.Due to the problems of various types of skin wounds,large differences in size,and blurred boundaries,it is a great challenge to realize the automatic segmentation of skin wounds.The main work and innovations of this paper are summarized as follows:A feature augment network(FA-Net)is proposed for the segmentation of skin wound images.In order to make full use of the edge information of the wound,a dual-stream feature extraction encoder model is proposed to extract segmentation features and edge features.The edge feature augment(EFA)module is added to the edge feature extraction branch.The attention mechanism is used to augment the extraction of edge features.An edge loss function is used to constrain network training.In order to exploit the spatial-relationship features between different wounds and between wounds and skin in the same image,the spatial-relationship feature augment(SFA)module is added between the network encoder and decoder to improve the segmentation accuracy of the network.An interactive feature augment network(IFA-Net)is proposed for the segmentation of skin wound images,thus forming a two-stage skin wound image segmentation method.In the first stage,initial automatic segmentation is obtained using FA-Net.In the second stage,user interaction is added to the initial segmentation to indicate segmentation errors,and the Gaussian distance is used to convert interaction tokens into interaction information that can be recognized by the network.Then the IFA-Net is constructed using FA-Net as the basic network structure,which takes user interaction information and preliminary segmentation results as input,and outputs optimized results.A software system for skin wound image analysis is also designed and developed based on PyQt,which realizes the functions of two-stage interactive skin wound segmentation,quantitative calculation of skin wound area and longest diameter,and printing of patient wound analysis reports.FA-Net and IFA-Net are evaluated on a skin wound dataset consisting of 552 images and achieves competitive results.The average Dice Similarity Coefficient(DSC)and average Intersection over Union(IoU)results of FA-Net are 89.16±0.79%and 82.16±1.06%,respectively,showing that the proposed automatic segmentation method can achieve decent wound segmentation.The DSC and IoU for IFA-Net are 94.03±0.87%and 89.65±1.14%,respectively,showing that the proposed interactive segmentation method can improve the segmentation accuracy based on simple manual markers.