Research On Fundus Image Segmentation Method Based On Encode-decode Architecture

The structure of retinal blood vessels in fundus is closely related to some diseases.Doctors can diagnose diseases such as glaucoma,diabetes and hypertension according to the morphological characteristics such as the diameter,size and curvature of retinal blood vessels.Accurate segmentation of retinal blood vessels is a key step in the diagnosis of diseases.However,at present,retinal blood vessel segmentation mainly relies on manual labeling,which is time-consuming and laborious,and easy to be affected by personal subjective factors.Retinal blood vessels present various tree-like structures with complex structures,which are easily affected by illumination,imaging equipment,diseases and other factors.It is difficult to achieve accurate segmentation of retinal blood vessels only by manual labeling.Therefore,it is necessary to design a computer automatic aided diagnosis system to accurately segment retinal blood vessels,so as to improve the diagnostic efficiency of doctors.For accurate retinal blood vessel segmentation,three algorithms are proposed in this paper: multi-scale attentional refinement retinal blood vessel segmentation algorithm,U-Net multi-scale self-calibration retinal blood vessel segmentation algorithm and double decoding path scalable dense convolution retinal blood vessel segmentation algorithm.Main research:(1)Considering that the traditional U-Net network has a simple structure and a single feature extraction,it is difficult to cope with the complex and variable characteristic structure of blood vessels.Based on U-Net structure,this paper proposes a multi-scale attention thinning retinal blood vessel segmentation algorithm.Firstly,an improved dense convolution module was used to extract the multi-scale feature information of blood vessels in the encoding and decoding stages.While reducing the number of parameters,the shallow feature information was fully utilized and fused to preserve more details.Secondly,a bidirectional attention mechanism is designed to analyze the importance of each pixel from the horizontal and vertical directions,and dynamically weigh the importance of blood vessels and background to eliminate the interference of background and noise.Finally,a spatial refinement structure was introduced in the decoding path to further extract the spatial structure of blood vessels,reduce background artifacts,and refine the shape of blood vessels.(2)In order to further improve the accuracy of blood vessel segmentation and improve the recognition ability of small blood vessels,a multi-scale self-calibrating retinal blood vessel segmentation algorithm is proposed which integrates scalable cascade module,Transformer and self-calibrating attention mechanism,etc.,to improve the effect of network segmentation.Firstly,a scalable cascade module is used to learn the complex and changeable retinal vascular topology in the coding stage and decoding stage,and channel attention is embedded to emphasize useful feature information and enhance feature transmission.Then,Transformer module is embedded at the bottom of the network to effectively capture the feature information of the global context of blood vessels,which is conducive to the spatial reconstruction of retinal blood vessel information.Finally,a self-calibrating attention mechanism is proposed in the jump connection stage,which adaptively calibrates the importance of features between the channel and space of the feature map,enhances the feature response of the target region,inhibits irrelevant features,effectively integrates shallow and deep features,and improves the discrimination ability of the network.(3)In order to improve the problem of vascular disconnection and further solve the problem of insufficient segmentation of small vessels,a scalable dense convolution vascular segmentation algorithm based on double decoding path is proposed.Firstly,a scalable dense convolutional module is designed to efficiently fuse shallow feature information,improve the transmission efficiency of small blood vessel information and accelerate the convergence of the network.Then,a feature activation module is introduced to effectively integrate the shallow and deep features through the learning channel and spatial interdependence,and improve the network’s attention to the target region.Secondly,a multi-level feature fusion module is introduced to expand the receptive field of network feature extraction and promote feature fusion at different scales.Finally,double decoding architecture and self-calibration module are introduced,and the output of encoder and the first decoder is further fused by the self-calibration module,and the output characteristics of the two paths are referenced to improve the vascular structure.The above algorithms were tested on DRIVE,STARE and CHASE＿DB1 public data sets.The experiments showed that the three algorithms proposed in this paper could segment as many small blood vessels as possible while keeping the main blood vessels continuous and intact,and compared with the algorithm of retinal blood vessel segmentation in recent years,the proposed algorithm was superior to most other algorithms.