Research On The Technology About Detection Of Glaucomatous Optic Cup And Disk Based On Deep Learning

Glaucoma is one of the two major blinding diseases in the world,and the number of patients is increasing year by year.However,the distribution of medical resource in our country is not balanced,the number of senior ophthalmologists is few and mostly concentrated in the central city,with few senior glaucoma professionals in remote areas.Therefore,with the aid of artificial intelligence technology,the research of computer aided automatic diagnosis technology for glaucoma is very necessary.CDR(the vertical height ratio of optic cup to optic disk)is one of the important reference index to diagnosis glaucoma,which can be calculated by both vertical height ratio of optic cup and optic disk after segmentation in the fundus.Nowadays the diagnosis research detecting of optic cup and optic disk of glaucoma based on computer aided technology mainly divided into two classes: one is to use digital image processing technology to detect the features of optic disk and optic cup on the fundus image to realize the orientation and segmentation of them,such as boundaries and texture;one is to using machine learning to realize the automatic positioning and segmentation of optic disc and optic cup.The above methods mainly have the following shortcomings:(1)Relying on the prior knowledge of the researcher,who need to have certain knowledge of glaucoma and image processing to extract relevant characteristics.(2)The sample data is small,the models is easy to overfit,the generalization ability is not strong,and the training speed is too slow.In the view of above shortcomings,this paper proposes a detection solution for glaucoma cup and disc based on generative adversarial networks.The main tasks include:(1)Date enhancement of fundus images can be realized by using techniques such as disk positioning,image scaling,cropping and rotation transformation,laying a good data foundation for later deep learning training.(2)An improved U-Net network structure,Light U-Net,is proposed to realized automatic detection and segmentation of cup and disk.(3)Combining the generative adversarial mechanism with Light U-Net,it realizes an automatic segmentation method of cup and dick based on generative adversarial network,Light U-GANs.The experimental results on public ophthalmic dataset ORIGA show that the average IOU of optic disc segmentation in this paper is 90.25%,while 70.88% for optic cup,the training time of the model reaches 0.0577 seconds/sheet,the test time reaches 0.0266 seconds/sheet,while training speed is 17 sheet/second,and the test speed is 38 sheet/second.