Retinal Optic Disc And Blood Vessel Segmentation Based On Deep Learning

The eye is not only an important window for observing the world,but also has an important responsibility to reflect the characteristics of the body.The characteristics of the retinal blood vessels at the bottom of the eye are the basis for judging various diseases.However,the structure of retinal blood vessels is complex and diverse,resulting in a difficult observation and a risk of missed detection.Therefore,the automated retinal vascular segmentation technology is of great significance for clinical diagnosis and has a wide range of practical value and application prospects.Although the retinal blood vessel segmentation methods are emerging,the traditional machine learning segmentation algorithm only uses a single feature to train the classifier,so the accuracy loss is severe and the segmentation effect is poor in practical applications.In recent years,deep learning has been widely used in various fields such as image processing.Therefore,in this dissertation,deep learning is introduced into the area of retinal disc and vessel segmentation.The main work of the paper is as follows:(1)Aiming at the problem that the retinal optic disc segmentation effect is susceptible to background pixels in the retinal image of the eye,two improved segmentation algorithms are proposed.The one is an improved U-net algorithmcombining attention mechanism and residual structure.The other is a generative adversarial network video disc segmentation algorithm,which uses the GAN game mechanism to fuse the deep separable convolution with the U-net structure to improve the generator structure.The discriminator uses convolution to generate the network structure.Both algorithms are verified on the DRIONS-DB dataset,and experimental results show that the accuracy of both algorithms is higher than current most of the deep learning algorithms,and the latter requires no positioning compared to the former,further improving the segmentation robustness and accuracy.(2)Aiming at the problem of low segmentation accuracy of the vascular end due to the complicated topological structure of the retinal vessels,a new improved neural network structure W-net(improved W-net)is proposed to improve the segmentation accuracy of the conditional generative adversarial network(CGAN).The improved W-net improves the generator in CGAN,resulting in an improved W-net-CGAN.Afterwards,the improved W-net-CGAN completes training and testing on the DRIVE dataset.Experimental results show that the improved W-net-CGAN has a higher segmentation accuracy than traditional retinal segmentation methods,as well as the fully connected network(FCN)and current most of the deep learning based semantic segmentation image segmentation network(including SegNet).The edges segmented are flatter,well guaranteeing the connectivity between blood vessels,especially on the segmentation of tiny blood vessels.In this dissertation,the DRIONS-DB and the DRIVE databases are used for testing the retinal disc segmentation and the vessel segmentation algorithms,respectively,on the Ubuntu platform using Python.The experimental results verify the feasibility and effectiveness of the proposed algorithms in this dissertation.