Research On Fundus Blood Vessel Segmentation Algorithm Based On Deep Learning

The retinal blood vessels present a continuous tree-like structure,and their morphological characteristics are an important basis for the diagnosis of ophthalmic-related diseases.Diabetes,hypertension,glaucoma and many other diseases can easily lead to fundus vascular disease.The application of retinal blood vessel segmentation technology to the quantitative analysis of ophthalmic disease symptoms plays an important role in clinical diagnosis.However,due to the complex and changeable morphological structure of retinal blood vessels,especially the capillaries in the fundus,the presence of uneven illumination and noise will make the task of blood vessel segmentation more difficult.Manual fundus blood vessel segmentation has the problems of low work efficiency and strong subjectivity,which is not only time-consuming and labor-intensive,but also very dependent on the professionalism of the ophthalmologist.The use of retinal blood vessel automatic segmentation technology to assist doctors in diagnosis can improve work efficiency and eliminate the influence of subjective factors,which has realistic medical diagnosis significance.However,small blood vessels are difficult to distinguish from the background,mis-segmentation of optic disc boundaries,hard exudates,and pathological spots are the problems faced by the task of blood vessel segmentation.At this stage,the fundus image segmentation model based on deep learning is limited by the network structure,resulting in insufficiently detailed segmentation of small blood vessels and prone to pathological missegmentation problems,and there are phenomena of missed detection and false detection.For this reason,based on the inherent characteristics of fundus images,this paper designs and establishes a suitable automatic segmentation framework for retinal blood vessels,which can effectively improve the accuracy of blood vessel segmentation.The main work is as follows:(1)Aiming at the problem that the actual receptive field of the network is far less than the theoretical receptive field of the existing fundus image segmentation model,the small blood vessel segmentation is insufficient and the multi-scale feature of the blood vessel cannot be captured,a multi-scale attention parse network is proposed.The parallel multi-branch structure and spatial pyramid pooling in the network can capture multi-scale features of blood vessels,increase the receptive field of the network and reduce the loss of contextual information between different sub-regions of the fundus image.The designed attention residual block strengthens the characteristics of blood vessels and suppresses background noise,so that the input layer close to the network can also obtain the global receptive field.In the network training,a booster training strategy is added to enhance the representation of blood vessel characteristics and increase the distinction between blood vessels and hard exudates.(2)Aiming at the problem that the U-Net-based network model cannot adaptively capture the multi-scale morphological features of blood vessels,and the semantic gap between the encoder and the decoder during the skip connection causes the loss of blood vessel details,adaptive attention residual network is proposed.Different from traditional multi-scale feature fusion technology,the dual attention module in this network can adaptively capture the multiscale morphological feature information of blood vessels,avoiding the interference of manual design of convolution kernel and subjective factors of pooling kernel.At the same time,the feature splicing and fusion module reduces the semantic gap between the encoder and the decoder,and can retain more detailed features of blood vessels.The designed soft attention residual block combines channel weight information and spatial position information,and the extracted blood vessel features are more abundant.The experimental results show that the retinal blood vessel segmentation method proposed in this paper has excellent blood vessel segmentation performance.It can segment more small blood vessels and is not easy to break,the thickness of the main blood vessels is appropriate,and it is closer to the real label.It still shows strong robustness and generalization in pathological images,and can effectively avoid the influence of objective factors such as uneven illumination,macular degeneration,and hard exudate.In terms of performance evaluation indicators,the sensitivity,F1 value and other key evaluation indicators of the method in this paper are higher,which is better than the method of comparative experiments,and has certain advantages compared with other existing advanced methods.