The Method Of Retinal Blood Vessels Segmentation In Fundus Image Based On Deep Learning

In clinical medicine,fundus images can reflect the early symptoms of various diseases,and ophthalmologists can achieve early diagnosis through retinal vessels.In the real world,the structure of retinal vessels is complex and doctors are subjective in diagnosis.Therefore,it is very important to design an automatic retinal vessel segmentation method to reduce the burden of doctors in early diagnosis.At present,deep learning has become the mainstream method in the field of medical image.Compared with the traditional segmentation method,it avoids the process of manual feature extraction and makes the final segmentation result more objective.Based on the above background,this paper uses deep learning to segment retinal vessels from the following three aspects:1.Segmentation method based on encode-decode network and attention mechanism.In this paper,a two-layer U-Net encode-decode structure is proposed,which is optimized by skip and bilinear interpolation and other methods to effectively avoid over fitting caused by small data sets.In order to capture the most favorable region in the feature map for segmentation task,this paper uses an attention module to extract the region of interest.The attention module makes full use of the vessel context information by recalibrating the weights in the channel and spatial domain of the feature map.Through the contrast experiment,the attention module is embedded in the appropriate position,and finally the high-precision vessel segmentation results are obtained.2.Segmentation method based on multi-scale and cascade network.In order to solve the common problem of disconnection in the task of blood vessel segmentation,this paper uses multi-scale information and cascade network to improve the related performance.Firstly,the multi-scale information of the feature map is captured based on the parallel branch and series branch structures in the multi-scale feature fusion method,and the specific implementation relies on dilated convolution and additional skip.The output of the main network is used as the input of the refine network,so that the refine network can correct the fracture of the blood vessel.In this paper,a three-path connection method is adopted,and the basic encode-decode networks with different multi-scale feature fusion structures are used as the main network and the refine network respectively.Finally,the vessel segmentation results with good connectivity are obtained.3.Application of uncertainty quantification in retinal vessel segmentation.Based on the segmentation and annotation images of multiple doctors in the dataset,this paper introduces the concept of uncertainty quantification in the vessel segmentation task,which aims to prevent the vessel segmentation model from making overconfident decisions.On the basis of bayesian neural network,the MC dropout method is used to obtain the image segmentation results of multiple forward propagation.After calculating the variance,the specific decomposition formula is used to decompose the variance into aleatoric uncertainty and epistemic uncertainty.The purpose is to simulate the labeled images in different situations and guide the direction for solving the shortcomings of vessel segmentation.