Research On Retinal Vessels Segmentation Method Based On Information Fusion

Diabetic retinopathy,cardiovascular disease,hypertension,arteriosclerosis and other diseases have different effects on retinal blood vessels.These diseases can be diagnosed by analyzing the characteristics of the blood vessel’s length,width,angle,curvature and branch in the retinal fundus images.Manual segmentation of the retinal blood vessels is a tedious,complicated and highly professional task,while the segmentation criteria is highly subjective.The computer-aided diagnosis system is employed to improve the diagnosis efficiency and reduce the misdiagnosis rate of doctors.It is significant to design an advanced segmentation algorithm to achieve rapid and automatic segmentation of retinal vessels.Therefore,two segmentation algorithms are proposed in this paper: retinal vessel segmentation algorithm based on multi-scale information fusion U-net and retinal vessel segmentation algorithm based on W-net conditional generative adversarial Nets.Its main work is as follows:(1)Retinal vessel segmentation algorithm based on multi-scale information fusion U-net.Considering that the traditional U-net model has the defects of single scale and poor information integration ability,the multi-scale blocks and attention mechanism are integrated into the U-net model.The multi-scale residual blocks are used to solve the single-scale network problem while avoiding gradient disappearance and gradient explosion.And the dense multi-scale dilated convolution blocks are employed at the bottom of U-net.It can not only expand the local receptive field,but also maximize the promotion of information fusion between different scale dilated convolution layers.In addition,the attention mechanism is used in the skip connection to effectively integrate the shallow features and deep features,and solve the problem of weight dispersion.(2)Retinal vessel segmentation based on W-net conditional generative adversarial nets.In view of existing segmentation algorithms are challenged with low sensitivity and insufficient segmentation of microvascular,a novel retinal vessel segmentation algorithm based on conditional generative adversarial nets using W-net as generator is proposed.The first is to expand the U-net to W-net through the skip connection.It is beneficial to the microvascular information transmission of the skip connection layer,accelerate the network convergence,and improve the parameter utilization.The standard convolutions are replaced with the depth-wise separable convolutions for expanding the network and reducing the number of parameters.The residual blocks are employed to mitigate the gradient disappearance and the gradient explosion.Each skip connection follows the Squeeze-and-Excitation blocks,and through learning the interdependence of feature channel,the shallow features and deep features can be effectively fused.Then,the loss function of conditional generative adversarial nets is modified to have strong global penalty ability in the game learning mode,and the overall segmentation performance is optimal.In this paper,the feasibility and effectiveness of the algorithms are verified on the DRIVE and STARE public datasets.Our algorithms can segment more tiny blood vessels as accurately as possible under the condition that the main blood vessels segmentation is continuous and complete,and the overall segmentation performance and sensitivity are better than most retinal vessel segmentation algorithms.