Research On Fundus Vessel Segmentation Algorithm Based On Base Location Features And Attention

The structure and morphology of ocular fundus vessels are the important basis for the diagnosis of various ocular diseases.Vascular segmentation is an indispensable work in clinical diagnosis.Using computer analysis method to achieve automatic segmentation of ocular fundus vessels can effectively improve work efficiency and save medical resources,which has very important practical significance.In recent years,deep learning has gradually become the mainstream method in the field of medical image segmentation.Due to its symmetrical up-sampling and downsampling process,U-Net network can obtain more accurate segmentation effect and is the most effective medical image segmentation model at present.However,in the segmentation task of fundus images,due to the small and uneven location distribution of the blood vessels in the retina,the location features are easy to be lost in the process of extracting blood vessel features from U-NET network,which leads to the decreased effect of restoring the location features in the up-sampling process.Based on this problem,this paper improves the SKIP process on the basis of U-NET network,so as to achieve better segmentation effect.The specific work contents are as follows:1.In view of the difficulty in obtaining the deep location features of U-NET network,the position enhancement module is added to U-NET algorithm and the LE U-NET algorithm is proposed.The position enhancement module uses void convolution to obtain the image multi-scale position feature information,and adds the feature to the original U-NET network sampling process.On the basis of not changing the original U-NET network feature extraction process,more feature information is added to the whole network.2.In the upsampling process of Le U-Net algorithm,three features are fused.In order to allocate the weight of the three features in the network fusion effectively and obtain more accurate segmentation results,attention module and fusion module are added to Le U-Net algorithm,and LEA U-Net algorithm is proposed.The attention module is integrated into the SKIP process so that the network can train the weight of attention,so as to adjust the weight ratio of the SKIP process and increase the proportion of the position strengthening module.The attention module and the position enhancement module are integrated into the upsampling process of U-NET using the fusion module.3.Make a comparative test of the model in the DRIVE data set.Firstly,in order to improve the image effect and highlight the image features,gray conversion,adaptive histogram equalization of limited contrast and gamma correction were used to preprocess the data,and the effects of several preprocessing methods on improving the accuracy of vascular segmentation were analyzed through experiments.Secondly,the two proposed algorithms are compared with the mainstream algorithms.By analyzing the experimental results,it is found that the segmentation effect of LE U-NET algorithm with position enhancement module is better than that of U-NET algorithm.After using attention module and fusion module to assign weight to the three features in LE U-NET,the accuracy of LEA U-NET algorithm has been further improved.