Research On Medical Image Segmentation Of Retinal Vessels Based On Image Semantic Segmentation

The level of human health is closely related to the development of medical science and technology.Intelligent medical treatment is an important way to improve the level of medical treatment in China,to solve the waste of medical resources and to reasonably allocate medical resources.It is also a new technology that has been vigorously developed in recent years in China.As the only human vascular network which can be directly observed noninvasively,the retinal vascular network of the fundus is of great significance to judge the health status of human body.Fundus retinal imaging can reflect the early symptoms of many eye diseases and systemic diseases,which is of great significance for the early diagnosis of diseases.Medical image segmentation can provide some reference for disease diagnosis,plays an important role in computer-aided diagnosis,and is also one of the hot research directions in recent years.This study focuses on the segmentation of retinal vessels.The main research contents are as follows:First of all,because of the poor illumination condition of the fundus retina image and the complexity of the tissue structure in the background,the image preprocessing method is used to make the display effect of the blood vessels clearer.The purpose of image preprocessing is to highlight the details of blood vessels,weaken the influence of background on blood vessel segmentation,and speed up the operation of neural network.The preprocessing process includes gray conversion,standardization,adaptive histogram equalization with limited contrast and gamma correction.Secondly,because of the difficulty of medical image acquisition and the high cost of manual annotation,the local image strategy is adopted.In the model training stage,multiple fixed size local images are taken from the complete image by randomly taking the center point,and the local image is used as a new training set for model training.In the test phase,the continuous and partially overlapped samples are taken on the test image at a fixed step through the mode of sliding window,and the sampled image is used for model test,and the overlapped part is used to improve the prediction effect by averaging multiple predictions.Then,the u-net network is improved based on the self-attention mechanism,and the global information module is constructed by the self-attention mechanism to solve the problem that the long-distance information can not be efficiently transmitted due to the local connection.By adjusting the output size of Q-transform,the global information module can be used in down sampling module and up sampling module.Finally,the feasibility of this method is demonstrated by experiments.The experimental results show that the AUC of this method is 0.9887,which can better segment the blood vessels from the background.Further analysis of the details of the segmentation results shows that this method has a certain adaptability in the interference of the highlight area,low contrast area and pathological exudate.