Research On Fundus Image Segmentation Algorithm Based On Deep Learning

With the development of the society,people’s life has changed.Nowadays,a variety of electronic intelligent devices can be seen everywhere,unhealthy schedules and excessive eye fatigue have led to more people suffering from fundus diseases.However,vision is the vital method for people to obtain information from the outside world.Once pathological changes occur,people’s normal life will be seriously affected.A variety of eye diseases will cause different degrees of changes to the fundus structure,fundus image is the most direct and effective basis for doctors to diagnose eye diseases,so fundus image analysis is of great significance for the auxiliary diagnosis of eye diseases.Early analysis of fundus images,including retinal vessel segmentation and optic disc detection,doctors can effectively identify these diseases based on the above images.The existing methods lack sufficient discrimination ability for fundus images,which are susceptible to lesions.Based on this,the following studies are conducted on fundus images in this paper:In order to solve the issue of microvascular segmentation,this paper proposed a U-shaped retinal blood vessel segmentation algorithm based on multi-scale attention feature aggregation network.The multi-scale residual dense convolution module is used to replace the original convolution layer to enhance its feature extraction capability.The improved attention mechanism module is added to make the network focus on blood vessel information.The dense dilated convolution module in the skip connection can further extract multi-scale information and mitigate the loss of feature information caused by down-sampling.Experimental results show that the proposed network can effectively solve the problem of microvascular segmentation and achieve accurate vascular segmentation.In order to solve the problems of vascular connectivity and vascular rupture,this paper proposes a retinal vascular segmentation network based on multi-scale feature fusion attention network.First,Drop Block is introduced into the residual dilated convolution module,which can effectively avoid overfitting.At the same time,the proposed residual multi-scale pooling layer can extract multi-scale context information.Multi-scale feature fusion attention mechanism can capture global context and semantic information.Experimental results show that,the segmentation network combine with the multi-scale feature fusion attention mechanism can preserve the vascular edge better,has better vascular connectivity,and effectively solves the rupture of the end of the main blood vessel.Aiming at the problem of the small number of fundus images and poor algorithm robustness,a generative adversarial network framework based on adaptive receptive field is proposed to segment retinal vessels and optic discs.The framework is composed of Generator and Discriminator,and the two games compete with each other until they reach equilibrium.In this paper,the multi-scale feature extraction module,the enhanced attention mechanism and the adaptive receptive field module are added to the generator,and the loss function of generative adversarial network is improved.The experimental results declare that the proposed framework can effectively solve the problems of small number of fundus images and poor robustness in the segmentation of fundus images.The fundus image segmentation network proposed in this paper can efficiently complete the segmentation task of retinal vessels and optic discs,and provide technical support for disease diagnosis and large-scale disease screening based on fundus image.At the same time,generative adversarial network provides a new way to solve the problem of training the dataset with fewer fundus images.