Research On Retinal Image Blood Vessel Segmentation Method Based On Convolutional Neural Network

Retinal blood vessel segmentation has important research value in the diagnosis of diabetic retinopathy,hypertension,cardiovascular and cerebrovascular diseases.Diabetes can induce many complications,among which ocular complications include diabetic retinopathy,cataracts,glaucoma,etc.,and their blindness is irreversible.The fundus is the only part of the human body where blood vessels can be directly observed.Its own changes,such as blood vessel width,angle,branch shape,etc.,provide a basis for early diagnosis of the disease.Fundus blood vessel analysis is currently the main way to diagnose fundus diseases,and fundus blood vessel segmentation is a necessary step for the quantitative analysis of diseases.Early analysis of retinal fundus images can effectively prevent such diseases.However,manual evaluation of retinal images by professional physicians is time-consuming and expensive,and is not suitable for screening large numbers of people.Although there are some automatic segmentation methods for retinal blood vessels,most of the early methods based on deep convolutional neural networks(DCNN)lack sufficient discrimination for fundus images,are easily affected by pathological areas,and have no large receptive fields.There is also no rich spatial information,and it is impossible to capture the global context information of a larger area,so it is difficult to identify the lesion area and the segmentation efficiency is low.In response to the above problems,this paper proposes a multi-scale and multi-path deep learning model based on convolutional neural networks.The main contents of the thesis are as follows:(1)A retinal blood vessel segmentation model based on a multi-scale dilated convolutional neural network is proposed.The model mainly contains three parts.First,an image preprocessing method based on automatic color enhancement(ACE)technology is studied to improve image quality,Strengthen the blood vessel area to achieve a better segmentation effect.Secondly,an improved deep fully convolutional neural network structure called BFCN is proposed for automatic segmentation of retinal blood vessels.Compared with the basic full convolutional neural network,BFCN has the following advantages: 1)Multi-scale input can effectively improve the quality of segmentation;2)Dilated convolution with different expansion rates is used to obtain larger receptive fields and rich spatial information.Fully understand the local context information;3)The conversion module uses the global average pool on the output of the encoding path to calculate the attention vector to guide the feature mapping learning.It can improve the network’s sensitivity to information features.In the absence of any monitoring information,information features are very important in the decoder,and effective encoder information can make better predictions;4)the sharpening method based on the Laplacian operator is studied to reprocess the predicted segmented image to repair the broken small blood vessels and achieve the purpose of improving the accuracy of retinal blood vessel segmentation.The experimental results show that the proposed multi-scale hole convolutional neural network model BFCN has better segmentation results than other models.(2)The proposed multi-scale dilated convolutional network segmentation of retinal blood vessels,compared with the previous deep convolutional neural network model,the model has a higher result,but because the basic convolution block is replaced with a multi-scale information extraction module,The network width increases,which will cause lengthy calculation time.Considering the above problems,this article continues to make further amendments to the model,hoping to reduce the calculation time without loss of accuracy.An effective semantic segmentation method is proposed,which combines convolutional neural network and recurrent neural network,and proposes a multi-path recurrent U＿Net network architecture to achieve segmentation of retinal medical images.First,multi-branch the encoding path and the decoding path to obtain different semantic features more effectively.Secondly,combined with the recurrent neural network,the multi-path output characteristics are time-sequential to further improve the target characteristics.Experimental results show that the multi-path cyclic U＿Net network model further improves the accuracy of retinal vessel segmentation.