Retinal Vascular Segmentation Algorithm Based On Multi-scale And Attention Mechanisms

Eye images play an important role in diagnosis and treatment of ophthalmological diseases(such as hypertension,arteriosclerosis,and diabetic retinopathy).Its retinal vascular morphological information can be used as an important indicator of these diseased diagnosis.Therefore,it is very important for the precision segmentation of the retinal vascular blood vessels,but due to the complexity of the retinal vascular structure and the different standards,artificially target the omentum blood vessels is a tedious and time-consuming task.With the help of the retinal vascular segmentation system,the retinal vascular vascular can be achieved.Ophthalmology expert work burden and improve diagnosis efficiency.To this end,this article conducts in-depth research on the intelligent segmentation of retinal vascular blood vessels,and proposes three retinal vascular segmentation algorithms: balanced the retinal vascular segmentation algorithm of the multi-scale attention network,the retinal vascular segmentation algorithm and cross-level fusion door of the multi-scale dense attention network Control the retinal segmentation algorithm that adapts to the network,and use the cross-level fusion door control self-adaptive network model to design a retinal vascular segmentation system with the cross-level fusion door control.The main research content is as follows:(1)Aiming at the different levels of the blood vessels of the omenture of the omenture in the eye image and the problem of regional interference in the lesion,a balanced multi-scale attention network is proposed to divide the retinal blood vessels.The network first uses a multi-scale feature extraction module encoding to obtain multi-scale context information to reduce the loss of vascular detail feature information;secondly,the details are introduced at the jump connection to enhance the module to increase the sensitivity of the network to the characteristic information and highlight the target area;then build a calibration.The residual module decodes to formulate noise and reduce overfitting.Finally,build a balanced scale feature module,integrate feature information at all levels,balance details characteristics and semantic characteristics,and reduce pseudo-shadow phenomena.(2)Considering that the artificial image data of the retinal vascular segmentation is limited,in order to more accurately divide the small blood vessels and improve the generalization of algorithm,a multi-scale dense attention network is proposed for the segmentation of the retinal vascular segmentation.The network constructs SCSE attention dense blocks instead of the traditional convolutional layer to coded and decoding,enhance the dual calibration of characteristic reuse capabilities and realize characteristic information;Improving the ability to obtain deep semantic characteristics of the network.(3)Around most algorithms,there are problems such as dividing vascular border blurry,fine blood vessels,and containing noise.In order to more accurately and quickly divide the retinal blood vessels,a lightweight cross-level door control self-adaptation network is proposed.The network uses a dense door control channel transformation module for encoding and decoding to promote competitive or synergistic relationships between channels to avoid the loss of shallow and thick-grained feature information;The upper and lower characteristics are avoided to avoid small blood vessels.In addition,the dual adaptive feature fusion module is used at the jump connection to guide the integration of the adjacent level characteristics to suppress noise.(4)In order to better auxiliary doctors clinical diagnosis of retinal vascular segmentation algorithms,a retinal vascular segmentation system was designed to engineering the results of the research results and serving the majority of eye retinal vascular segmentation needs.The system enters through the user login interface.The system contains four modules: import weight,loading image,vascular segmentation,and saving results.The interface is simple and simple.Among them,the vascular segmentation module uses the retinal vascular segmentation algorithm proposed in this article for division.This article is verified on the open dataset.As a result,the three retinal vascular segmentation algorithms mentioned in this article can better divide the thick and thin blood vessels with a complete structure,and the performance is better than the existing majority algorithm.The accuracy and real-time nature of the retinal vascular segmentation system also meet the requirements,and the operation is convenient and good.