Research On Fundus Segmentation Method Of Fundus Image Based On Deep Learning

There are many kinds of fundus diseases,such as high incidence of cataract,glaucoma,age-related macular degeneration,diabetic retinopathy,are the four leading cause of blindness,which do great harm to human body.Meanwhile,the ocular fundus is the only part of human body that can directly observe the blood vessels,fundus disease can also reflect other diseases of the body,such as diabetic retinopathy.Fundus image is the main way to diagnose these diseases.And the retinal image vessel segmentation is a necessary step for the quantitative analysis of the disease.At present,the best retinal image vessel segmentation methods are supervised ones,these methods require artificial design features which are heavily dependent on the designer's experience,meanwhile the optimization of the algorithm parameters needs a lo t of time.In recent year,thanks to the ability to automatically extract features and optimize algorithm parameters,deep learning has made a breakthrough in image processing,speech recognition and other aspects.Therefore,this thesis focuses on the app lication of deep learning in the retinal image vessel segmentation.In this thesis,we study two classes of retinal image vessel segmentation methods based on deep learning,which are patches-based image segmentation and end-to-end image segmentation.These methods all have strong ability of feature extraction,and their model can optimize their own parameters in the training process.The patches-based image segmentation methods output the corresponding label probability map when taking image patch as input for the first time;the end-to-end image segmentation methods are achieved for the first time,and they output the corresponding overall label probability map when taking an original retinal image as input.The four retinal image vessel segmentation methods proposed in the paper can automatically extract feature and optimize algorithm parameter,and the accuracy and speed of the algorithm are improved compared with the previous methods.O n DRIVE,STARE and CHASE,these three fundus image databases,the sensitivity(Se),specificity(Sp)and accuracy(Acc)of each method have reached the world advanced level.Among which,the end-to-end image segmentation methods have expert level performance on DRIVE and CHASE database,meanwhile these methods can significantly improve the vessel segmentation speed.In the current laboratory hardware condition,we use the quarter network model to segment a retinal image in DRIVE database that takes only 0.25 s.From the patches-based image segmentation method to the end-to-end image segmentation method,a variety of retinal image vessel segmentation methods are proposed from different aspects.These proposed image segmentation methods have the advantages of excellent performance and convenient application,and these methods have important value and significance to the whole field of image segmentation.