Improved U-shaped Network For Fundus Image Blood Vessel Segmentation Algorithm

The detection and analysis of vascular morphology in fundus images is an important part of the screening,diagnosis,treatment and evaluation of ophthalmic diseases.By analyzing morphological changes such as vessel length,width,branching patterns and angles,clinicians can accurately diagnose ophthalmic diseases.Vessel segmentation in fundus images is a key indicator for assessing the degree of ophthalmic diseases,therefore,the development of effective automatic vessel segmentation methods is crucial for the diagnosis of related diseases.Traditional manual segmentation methods are tedious,error-prone,and time-consuming.At the same time,the topological structure of blood vessels in fundus images is irregular,complex in shape,and varied in scale.It is difficult to accurately segment blood vessels in fundus images only by manual methods.Computerized segmentation strategies can achieve accurate segmentation of blood vessels in fundus images,reduce the annotation time and workload of ophthalmologists,and improve the efficiency of clinical diagnosis related to ophthalmic diseases.Therefore,this paper proposes three improved U-shaped networks for fundus image vessel segmentation algorithms,including Ushaped segmentation algorithms that incorporate multi-layer spatial attention,multi-level adaptive scale U-shaped segmentation algorithms,and multi-resolution fused input U-shaped segmentation algorithms.The main research contents are as follows:(1)Considering the insufficient segmentation of small blood vessels and weak anti-noise ability in fundus images,a U-shaped retinal blood vessel segmentation algorithm that combines multi-layer spatial attention is proposed.First,in the U-shaped network architecture,the traditional convolution is replaced by a feature enhancement residual module,and the channel attention mechanism is added to adaptively enhance the characteristics of the blood vessel channel;secondly,a dense hole convolution module is added at the bottom layer of the network to expand the feature receptive field and extraction Multi-scale features;Finally,a three-terminal spatial attention module is added in the skip connection stage to perform adaptive refinement on feature extraction while effectively suppressing noise interference in the fundus feature map.(2)Focusing on the characteristics of small blood vessels and complex scale changes in fundus images,a multi-level adaptive scale U-shaped retinal blood vessel segmentation algorithm is proposed.One is to introduce a residual module in the encoding and decoding structure,which can not only strengthen the channel feature propagation ability,but also strengthen the model’s organic selection of blood vessel semantics.The second is to embed a multi-scale feature extraction module at the bottom of the network to effectively capture multi-scale vascular features.The third is to build an improved adaptive feature fusion module between the encoding-decoding architecture,aiming to effectively fuse the features of the adjacent layers of the encoding to the decoding part,and extract more effective features of small blood vessels.The fourth is to add a multi-level attention structure in the decoding part to achieve adaptive refinement of multi-level feature information.(3)Aiming at the problems of low contrast between blood vessels and background areas in fundus images,large scale changes,and hard exudates from bleeding,a U-shaped network with multi-resolution fusion input is proposed to realize the detection of blood vessels in fundus images.Precise segmentation.Design a coarse network with a multi-resolution fusion input architecture as the backbone,maintaining high-resolution information features.The traditional convolution is replaced by the improved Res Ne St to optimize the boundary features of blood vessel segmentation;at the same time,a parallel spatial activation module is embedded to capture more spatial and semantic feature information.Another U-shaped fine network is constructed to enhance the microscopic recognition and representation capabilities of the model.One is to increase the multiscale dense feature pyramid module at the bottom layer to extract multi-scale feature information of blood vessels.The second is to use the feature adaptive module to enhance the feature fusion between coarse and fine networks to eliminate irrelevant background noise in images.The third is to integrate and design multiple loss functions for vascular texture details to enhance the network’s attention to vascular features.In this paper,the above three algorithms are verified on three public fundus datasets.The results show that the three segmentation algorithms can accurately segment small and complete vascular structures,and the overall performance is better than most existing algorithms.