Research On Retinal Vascular Segmentation Algorithm Based On Deep Learning

Retinal fundus blood vessels are located at the end of the human blood circulation system and are the only part of the body that can be observed vascular structure in a non-invasive way.The retinal vascular tree is unique to each individual and can be used for biometric identification.The fundus vascular images also contain rich characteristic information,which can be used for further diagnosis and treatment of ocular diseases,such as diabetes,hypertension,and coronary heart disease.The segmentation of fundus blood vessels is the most important step in the diagnosis of these diseases.Therefore,accurate segmentation of retinal blood vessels in the fundus can help doctors better diagnose related diseases,which is of great significance for providing appropriate treatment plans in the later stage.Due to the influence of uneven illumination,the quality of medical imaging systems,and diseases,it is difficult to segment the blood vessels from the background.The solutions proposed by researchers can be divided into manual segmentation,unsupervised method,and supervised method.Manual segmentation is costly and error-prone.Methods based on unsupervised segmentation often cannot obtain very accurate segmentation results.Therefore,this paper mainly adopts deep learn-based methods to achieve accurate segmentation of retinal vessels in the fundus.The research in this paper mainly includes the following aspects:(1)Retinal vascular segmentation algorithm based on GBU-NET.Firstly,data enhancement is performed on the original data,and the original image is divided into patch blocks.Then,the patch blocks are sent into the network model,and the patch blocks that run out of the model are combined with the segmentation results of the corresponding image.The network model is mainly improved based on U-NET.In the first three layers of downsampling,the global convolutional network combined with boundary refinement module was used to replace the common coding operation,to improve the model’s ability to extract information contained in different transformation feature maps and vessel boundaries.Dense Net module is used in the last layer of downsampling to prevent over-fitting of the model.(2)AGBU-Net vascular segmentation algorithm with dual attention mechanism is introduced.The network model is mainly improved based on GBU-NET.To modify traditional jump connection,by improving the position of the first note on better extraction module makes model sampling feature maps contain semantic information,and using the improved space look that module model better extraction and corresponding coding layer characteristic graph contains spatial information,and then use the Conv LSTM better combined with the feature of them to obtain information;The original fundus images were preprocessed to improve the contrast between blood vessels and background.Post-processing was used to reduce the noise points of the vascular probability distribution map obtained by the model.(3)A new vascular segmentation algorithm MDASPP-Net is proposed.The network model is an improvement of AGBU-NET.By inputting original images of different scales in the first three layers of the coding layer,different proportions of image feature information can be obtained.In the last layer of coding,a multi-scale dense feature pyramid cascade module is used to perform multi-scale dilated convolution cascade operations on feature images of different sizes to obtain local information and context information of feature images of different sizes.Then the model parameters are reduced and the model training speed is accelerated by replacing traditional convolution with deeply separable convolution.The sensitivity,specificity,accuracy,F1-score and AUC were evaluated in two public retinal image datasets DRIVE and CHASE＿DB1.The results showed that MDASP-NET performed better than other algorithms in vascular segmentation.