Research On Eye Image Segmentation Method Based On Improved U-Net

The structure and morphology of retinal vessels contain a lot of information,which can detect early symptoms of many diseases.Therefore,in the screening and diagnosis of eye diseases,the examination of retinal vessels is a very important component and is of great significance for the diagnosis and treatment of patients.However,due to the complexity and variability of retinal vessels,manual segmentation of vessels requires enormous human resources and time.In recent years,relying on a computer system for accurate and efficient automatic segmentation of retinal vessels has gradually become a development trend.This not only reduces labor costs,but also improves clinical diagnostic efficiency and has very good application prospects.For the task of retinal vessel image segmentation,this paper proposes three retinal vessel segmentation algorithms based on the mainstream segmentation medical image U-Net network,namely the improved residual U-shaped network algorithm,the improved dense Ushaped network algorithm,and the lightweight segmentation network algorithm.The main research work is as follows:(1)In response to the complexity of retinal blood vessel structure and the difficulty and inaccuracy of segmentation,a retinal blood vessel segmentation algorithm based on an improved residual U-shaped network is proposed.This algorithm focuses on the depth of the network.Firstly,a residual structure is introduced to form residual convolution,which is used to build the encoder and decoder based on the U-Net network.This can not only deepen the network to transmit more feature information and improve segmentation performance,but also avoid the problem of network degradation.Then,a spatial attention module is added to the bottom of the network and connected with the residual convolution to extract deep-level features.Finally,a channel attention module is added at the junction of the encoder-decoder feature connections to suppress noise and enhance key information.From the segmentation results,this method improves the accuracy and precision of segmentation and successfully segments the complete blood vessel image.(2)To achieve more accurate segmentation of retinal vessels and address the issues of small vessels appearing blurry and vessel branches being discontinuous in retinal images,an improved dense U-net networkbased retinal vessel segmentation algorithm was studied.Firstly,to fully utilize the feature information between convolutional layers,the advantages of dense networks and residual structures were combined.A dense connection residual module was proposed and used to construct the encoder and decoder of the network.Secondly,to extract rich feature information from the downsampling process of the encoder,dilated convolutions were introduced and a multi-feature distillation module was designed to extract features with different receptive fields and concatenate them before passing them into the decoder.Finally,to fully integrate lowlevel and high-level features,a bidirectional convolutional LSTM module was introduced at the skip connection to generate more accurate prediction maps.Experimental segmentation results showed that this method further improved the details of vessel features,reduced the discontinuity of vessel branches in the segmentation results and provided clearer outlines of small vessels.(3)In order to further improve the segmentation performance while reducing the number of parameters in the entire network model,a lightweight network-based retinal vessel segmentation algorithm was studied.Firstly,based on the U-Net,the HRNet network framework was introduced and three different resolution branches were designed to ensure that the network maintains high-resolution representation throughout.Then,ghost convolution was introduced to replace ordinary convolutional layers to reduce the number of network parameters.At the bottom of the network,a multi-scale extraction module constructed by dilated convolutions was added to extract multi-scale features of retinal vessels and enrich information.Finally,attention modules were embedded in different resolution branches to suppress noise,enhance detail features,and improve network segmentation performance.Experimental results show that the performance of the lightweight network is significantly improved,the number of parameters is reduced,the burden on hardware equipment is reduced and training efficiency is accelerated.