Research On Retinal Vessel Segmentation Method Based On Deep Learning

Retinal blood vessels are an important part of the human microcirculation system.Cardiovascular diseases such as diabetes and hypertension can cause changes in the morphology of blood vessels.Doctors can use images of fundus blood vessels as a basis to determine the type and severity of the patient’s disease.Therefore,the extraction of retinal blood vessels from color fundus images plays an important role in the prevention and analysis of cardiovascular diseases.The artificial retinal blood vessel segmentation method is time-consuming and expensive,not only requires a large number of ophthalmologists,but also cannot be screened nationwide,so it needs to rely on automatic segmentation technology.At this stage,a large number of automatic segmentation methods have been proposed.However,the retinal blood vessels are extremely complex.Not only have high curvature but also many capillaries,and there are lesion interference.In addition,the fundus camera has certain limitations in capturing retinal images,which makes the extraction of retinal blood vessels very challenging.Sex.Therefore,in order to explore and solve the above-mentioned problems,this paper makes the following research:(1)In order to deal with the influence of the inherent characteristics of retinal images on segmentation,this paper adopts preprocessing strategies to solve such problems,and discusses the different effects of the Retinex algorithm of channel weighting,normalization,CLAHE,Gamma correction and bilateral filtering on image enhancement.From the perspective of the preprocessing effect,combining these methods can effectively improve the gradient of the blood vessel edge and eliminate part of the noise.From the perspective of segmentation results,preprocessing methods can improve the evaluation results of the algorithm.(2)Aiming at the complex morphology and structure of blood vessels and rich semantic information,a U-shaped retinal blood vessel segmentation algorithm with multi-feature fusion is proposed.First,the images with enhanced blood vessels and increased brightness are obtained through preprocessing,and then the cropped images are input into the U-shaped network for training.The network uses dense convolution blocks to achieve feature multiplexing;adds SE and AGs attention modules to suppress the extraction of background information;and designs a multi-channel feature distillation module for the bottom layer of the U-shaped network to extract multi-scale Vessel semantic information.Experimental results show that the evaluation indicators of the algorithm are better than the U-net algorithm.(3)Taking into account the problems of the loss of small blood vessel details and the misjudgment of lesion information in the existing algorithms,a blood vessel segmentation algorithm based on improved HRNet is proposed.The algorithm inherits the advantages of the high information exchange rate of the HRNet,uses deformable convolution to adaptively enhance the network’s sensitivity to blood vessel pixels,and the designed long-distance attention module and multi-scale feature extraction module alleviate the problem of mis-segmentation,Improve the stability of the network.Experiments show that HRNet can segment more capillaries compared to U-net,but there are more mis-segmented areas;and the improved HRNet not only preserves capillaries,but also greatly reduces the misjudgment of background pixels.