| In recent years,deep learning has developed rapidly in the medical field and achieved doctor-level accuracy in many diagnostic tasks,even surpassing doctors,especially in the diagnosis,segmentation,and detection tasks related to medical images.Retinal vessels are the only tiny vessels that can be directly observed in the human body.Under normal circumstances,the morphology and structure of retinal vessels at the ground of the eye remain relatively stable,but some diseases can cause changes in the morphology of retinal vessels,such as diabetes and hypertension.The use of retinal vessel segmentation images is helpful to the prevention,treatment of vascular diseases and ophthalmic diseases,retinal vessel segmentation algorithm plays an important role in the basis of auxiliary diagnosis in the screening system of related medical diseases,people begin to pay attention to how to segment medical images quickly and accurately by the algorithm.Although the current research methods based on deep learning have achieved excellent results,there are still several problems in the retinal vessel segmentation task.(1)For the small blood vessels in the retinal vessels,there is still a problem of low segmentation accuracy,(2)The network model of the existing method is universal and has a huge amount of parameters.In this paper,we have conducted a base on deeping learning on retinal vessel segmentation methods and proposed three effective retinal vessel segmentation methods.First of all,to solve the discontinuity problem of retinal vessel segmentation,this paper proposes a retinal vessel segmentation network inspired by ordinary differential equations.From the point of view of numerical analysis,an ordinary differential equation module is designed.To further strengthen the ability of the module to obtain multi-scale information,the expansion convolution with different expansion rates is added to different branches of the module,which can effectively extract different scales of retinal blood vessels.Characteristics.Finally,to verify the effectiveness of the proposed network,this paper builds an experimental environment and conducts sufficient experimental verification.The experimental results show that the sensitivity of the proposed method is greatly improved on the CHASE＿DB1 dataset,and the ability of ODE-UNet to capture and classify blood vessel pixels is improved.Secondly,to improve the contextual information of retinal vessels,this paper applies the attention mechanism to the task of retinal vessel segmentation.Taking advantage of the feature of the automatic attention area of the attention mechanism,an end-to-end retinal blood vessel segmentation network based on enhanced spatial attention is designed.The network introduces a new enhanced spatial attention(Enhanced Spatial Attention,ESA)module.Compared with the spatial attention(SA)module,ESA can obtain a larger receptive field,to better adapt to the spatial context content.Redistribute features locally,and adding ESA modules can allow the network model to learn more powerful discriminative features.Experimental results show that the network model proposed in this paper has achieved better performance in terms of performance indicators and vision.Finally,to solve the problem of huge parameters,this paper proposes an end-to-end retinal vessel segmentation network(LMSFF-Net)based on lightweight multi-scale feature fusion.From the perspective of reducing the number of model parameters,the network adopts a strategy of reducing the depth and width of the network(that is,reducing the number of downsampling of the encoder and the number of channels in the convolutional layer).The network first adds short-hop connections to the encoder,and uses multi-scale feature fusion to aggregate feature maps of different scales and sizes;to further reduce the amount of parameters,the network designs a compression module to reduce the amount of parameters by compressing the number of channels.Finally,a comparative experiment was carried out.The experimental results show that the network proposed in this paper can effectively reduce the parameters of the model while obtaining better segmentation results. |
