Research On Medical Image Segmentation Based On Graph Convolution

In modern medical diagnosis,doctors usually diagnose the condition by analyzing medical images.For example: Breast cancer is the most common malignancy in women worldwide.For breast cancer detection,mammography is one of the most straightforward methods.In clinical practice,doctors can determine whether there is an abnormality in the breast by looking at X-rays.However,due to the complex structure of the breast and the lack of obvious lesions in the early stage of breast cancer,it is easy for doctors to misdiagnose.On the other hand,the number of patients in China is large,which requires doctors to diagnose the disease in a short period of time,which brings greater challenges to doctors' accurate diagnosis.With the development of computeraided diagnosis,medical segmentation systems based on deep learning have been applied in major hospitals to assist doctors in diagnosing diseases and improve the accuracy of medical diagnosis.In the past few years,traditional image segmentation methods have been gradually replaced due to the introduction of fully convolutional networks.For natural images,the accuracy of image segmentation based on deep learning has been greatly improved,but it has developed slowly in the field of medical image segmentation.The main reason is that medical images are different from natural images,and have the characteristics of unbalanced noise distribution and complex lesion tissue to be segmented.When using a fully convolutional network to obtain advanced features,due to the limitation of the receptive field,the local location information of some diseased tissues is not fully extracted,resulting in a decrease in segmentation accuracy.At the same time,the shortage of medical image labeling data also imposes great restrictions on network training.In order to solve the above problem.In this paper,the related exploration and improvement of unsupervised learning and supervised learning medical image segmentation algorithms are carried out.The main research contents are as follows:For supervised learning medical image segmentation,this paper uses graph convolutional networks combined with fully convolutional neural networks to extract high-level and low-level features of images.First,a fully convolutional network is used to extract the low-level features and some high-level features of the image.Then,the triplet information of the graph is constructed using the extracted features.Finally,highlevel features of the image are trained using a graph convolutional network.For unsupervised learning research,this paper adopts a hyperbolic graph convolutional network to fuse a domain adaptive image segmentation framework based on adversarial learning.By using a hyperbolic graph convolutional network,on the one hand,the distortion in the feature mapping process is reduced,and on the other hand,the structural distribution difference between the source domain and the target domain can be trained.In order to verify the effect of the research content in this paper,this paper conducts relevant experiments on the two methods respectively.The experiments used two public breast datasets: the Digital Database for Screening Mammography(DDSM)and the INbreast dataset.Compared with other commonly used frameworks for image segmentation,the results show that the method proposed in this paper is higher than other frameworks in several indicators such as segmentation accuracy,which reflects the effectiveness of the algorithm in this paper.