Research On Retinal Blood Vessel Segmentation Algorithm Based On Improved MNet And PCNN

According to the current medical research,retinal lesion is closely related to a variety of human diseases,such as diabetes,glaucoma,liver cirrhosis,nephritis,ophthalmic diseases and so on.In clinical medicine,retinal vessel segmentation is the basis of medical diagnosis,it is usually used to diagnose and screen the body for a wide range of associated diseases through lesions in the retinal vessels,and plays an important role in the screening and diagnosis of related diseases.However,the vascular structure of the retina is complex and diverse,and the vascular endings are different and tiny,making the segmentation of retinal vessels in the fundus purely dependent on manual segmentation and marking of lesions by the ophthalmologist.This not only increases the workload of doctors,but also tests the ability of doctors’ clinical practice and experience.At the same time,manual segmentation is greatly affected by the subjective factors of the doctor,so pure manual retinal vessel segmentation is likely to have the risk of underegmentation and missegmentation.Therefore,the use of new technological methods to segment retinal vessels is important for reducing physician workload,saving medical costs and improving the accuracy of medical diagnosis.Aiming at the problems of low accuracy and poor connectivity of the existing segmentation methods for blood vessels,this thesis takes the fundus image as the research object,and uses the relevant knowledge of digital image processing and artificial intelligence algorithms to study the segmentation of retinal blood vessels in the fundus.The main research work and results of this thesis are as follows.Firstly,preprocessing operations of retinal images,highlighting the features of blood vessels.The retinal image is subjected to normalization contrast-limited adaptive histogram equalization(CLAHE),and gamma correction processing to improve the contrast between the retinal vessels and the background,which is beneficial for subsequent segmentation of the retinal vessels.Secondly,construction and improvement of the neural network model to achieve coarse segmentation of retinal images.The encoder-decoder is used as the basic architecture of the network,and a multi-scale attention residual module with deformable convolution is introduced to make the receptive field of the convolution kernel closer to the actual shape of the object,which can target the capture of features in the region of interest of the feature map and assign different weights to the extracted feature information to improve the discriminative ability of the feature information.Thirdly,using the improved Pulse-Coupled Neural Network(PCNN)with dynamic multi-threshold segmentation instead of traditional single-threshold segmentation to achieve fine segmentation of retinal images.The neighborhood connection weight matrix of PCNN is improved and proposes a two-dimensional Gaussian normal distribution-based PCNN connection weight matrix,and the genetic algorithm is used to find the optimal parameter values of the PCNN model.The dynamic multi-threshold segmentation characteristic of the PCNN model is introduced into the neural network model,replacing the traditional single-threshold segmentation,making it easier for the network to obtain the global optimum segmentation.