Research On The Semantic Segmentation Algorithm Of Fundus Retinal Vessels

Retinal vessels in the fundus image serve as deep microvessels in human body,and can be directly observed.They have medical diagnostic value when deforming into other different shapes such as changes in diameter,length,branching,contortions,etc.Therefore,retinal vessels can be used for the detection,evaluation,and treatment of cardiovascular and cerebrovascular diseases and ophthalmic diseases.On the other hand,the acquisition of fundus images is convenient and non-invasive,so it is also very easy to observe the retinal vessel morphology.Automatic segmentation of retinal blood vessels can reduce the workload of ophthalmologists while improving the accuracy of segmentation,and can perform large-scale screening and analysis of ophthalmic and cardiovascular diseases coupled with computer assistance.Segmenting vessels from the retinal image is the first step when analyzing fundus images.However,it is quite difficult due to the complicated branch structure of the fundus vessels,the noisy background,and the difference in various lighting.This paper proposes the construction of a semantic segmentation network using deep learning algorithms to segment retinal vessels of the fundus image.In recent years,applying convolutional neural networks into image processing has become a research hotspot,especially into the fields of image segmentation,recognition,and detection.Image semantic segmentation is an important section in the field of artificial intelligence,deep learning,and machine vision.The semantic segmentation of retinal vessel,in short,is to classify each pixel of the picture into vessel pixels and non-vessel pixels.This paper constructs two different semantic segmentation networks which are RV-SegNet and RV-LinkNet,to perform retinal vessel segmentation.RV-SegNet(Retinal Vessel SegNet)is an improved semantic segmentation network based on SegNet for extracting retinal vessels.First,the dataset is preprocessed,and then the image is cut into small-sized images in an overlapping manner.Image pixel-to-pixelsegmentation is performed by the Encoder-multi-Decoder structure.Then,how Encoderblocks and Decoder-blocks with different receptive fields influence the segmentation effect is studied.Among those segmentation network structures,the one using the two layers of Encoder-block u and the one layer of Decoder-block achieves the best effect.It can be inferred that it is not rational to blindly increase the number of convolutional layers,but to make corresponding adjustments according to the receptive field and image size.Comparing the segmentation results of RV-SegNet and other methods on the DRIVE and STARE databases,RV-SegNet has higher accuracy,specificity,and AUC than other segmentation networks,indicating that the RV-SegNet method has high preciseness.RV-LinkNet(Retinal Vessel LinkNet)is an improved network based on LinkNet.First,the dataset is preprocessed and the image data is augmented.Then,the LinkNet is modified into RV-LinkNet,based on the conclusion about the relationship between the receptive field and image size obtained in the previous experiment.Lastly,Endoder modules and Decoder modules with dilated convolution and residual structure are designed to obtain a larger receptive field,without losing feature information and without increasing the amount of calculation.After series trainings and experiments,the best RV-LinkNet architecture is obtained,which has different advantages in accuracy(Acc),sensitivity(Sen),specificity(Spe)and AUC on DRIVE,STARE and CHASE_DB1 databases,and it outperforms other segmentation algorithms overallThe semantic segmentation networks proposed in this paper can efficiently complete the fundus retinal vessels semantic segmentation task and can provide technical support for ophthalmic disease diagnosis and large-scale disease screening based on fundus images.At the same time,the method proposed in the network is suitable for training on datasets with a small number of images,and provides new ideas for deep learning.