Research On Dermoscopic Image Segmentation Method Based On Convolutional Neural Network And Graph Neural Network

Skin cancer has a short outbreak period and is poorly cured in the middle to late stages.Patients need to be detected and surgically removed as early as possible before the cancer cells spread.However,the complexity of the lesion area and the increasing number of patients in the clinical diagnosis process pose serious challenges for physicians.Therefore,the use of computer-aided diagnostic techniques to enhance the accurate detection and segmentation of lesion areas in dermoscopic images and provide better diagnostic assistance to dermatologists has become the key to the problem.Currently,some segmentation methods based on thresholding,clustering algorithms and neural networks have been used to segment the edges of lesions in dermoscopic images.However,these methods suffer from the problems of not being able to take into account the local and global information contained in the dermoscopic images and insufficient learning of key features of the images.To address these problems,this paper combines convolutional neural networks and graph neural networks.In this paper,we combine convolutional neural networks and graph neural networks to perform the dermoscopic image segmentation task and validate the effectiveness of the method on colonoscopic and breast cancer ultrasound datasets.The main research contents of this paper are as follows:(1)We propose a stacked convolutional neural network incorporating a novel and efficient pyramid residual attention(PRA)module for the task of automatic segmentation of dermoscopic images.The proposed PRA has the following characteristics: First,we concentrate on three widely used modules in the PRA.The purpose of the pyramid structure is to extract the feature information of the lesion area at different scales,the residual means is aimed to ensure the efficiency of model training,and the attention mechanism is used to screen effective features maps.Thanks to the PRA,our network can still obtain precise boundary information that distinguishes healthy skin from diseased areas for the blurred lesion areas.Secondly,the proposed PRA can increase the segmentation ability of a single module for lesion regions through efficient stacking.The third,we incorporate the idea of encoderdecoder into the architecture of the overall network.Compared with the traditional networks,we divide the segmentation procedure into three levels and construct the pyramid residual attention network(PRAN).The shallow layer mainly processes spatial information,the middle layer refines both spatial and semantic information,and the deep layer intensively learns semantic information.The basic module of PRAN is PRA,which is enough to ensure the efficiency of the three-layer architecture network.(2)We build a pixel-to-pixel segmentation model called Graph reasoning and Inception Attention Network(GIAN).First,we propose a graph reasoning module that is data-dependent.The node matrix of the graph reasoning derives from the original image and the feature map,so our graph reasoning module can accurately capture global information in feature maps.Second,to avoid the information redundancy caused by channels,we propose the Inception attention module based on the original Inception module,which extracts the local spatial semantic information in the feature information.The Inception attention module can select representative node graphs as feature guidance graphs for image segmentation.The spatial information extracted by multiple parallel convolution kernels ensures the stability of subsequent pixel classification.In this way,the GIAN takes the extraction of global and the guidance of local information into account.The organic combination of the two modules provides a correct ideological basis for the segmentation task.In addition,we initially explored the generation of a graph reasoning multi-scale attention network model for antagonism enhancement,and improved the generalization ability of the model by using the distribution characteristics of the generated antagonism pattern learning data.(3)We evaluated the performance of the Pyramid Residual Attention Network and the Graph Reasoning Inception Attention Network on each performance metric on the ISIC2017 and ISIC2018 datasets.We analyzed the respective advantages and shortcomings of the two methods through comparing the experimental results.The experimental results demonstrated that the two models achieved 93.39%,87.94%,94.98%,95.56%,96.15% and 94.02%,88.98%,95.22%,95.96%,96.72% in Dice,MIOU,specificity,accuracy,and sensitivity on ISIC2018,respectively.On the ISIC2017 dataset,89.47%,82.14%,93.50%,93.17%,93.86% and 90.02%,83.08%,93.60%,92.80%,94.45% were achieved,respectively.They achieved good segmentation results compared to existing methods.