Diagnosis And Analysis Of Glaucoma Based On CNN

Glaucoma is a chronie eye disease in which the optic nerve is gradually damaged.It is the second most common blind eye diseases after cataract.The corresponding visual impairment is irreversible.The timely diagnosis of glaucoma at early stage is the key to treating and controlling the course of the disease.The incidence of glaucoma is concentrated in the optic nerve head area,mainly manifested by elevated intraocular pressure,optic atrophy,collapse and visual field defects.In the diagnosis of glaucoma,in addition to visual field defects and elevated intraocular pressure,it is easier to obtain and quantify the retinal cup-to-disk ratio.The glaucoma treatment joint research system has also allowed the cup-to-disk ratio to be registered as a standard forjudging glaucoma.Recently,most researchers are calculating cup-to-disk ratios by using digital fundus image.The calculation of cup-to-disk ratio is relatively single and relies too much on retinal cup and disk segmentation results.This paper proposes a method for detecting the position of the opening of the Bruch membrane opening on the OCT image,segmenting the inner limiting membrane of the retina,and calculating the retinal cup-to-disk ratio according to the position relationship between the two.The main work and innovations are summarized as follows:1.A convolutional neural network for detecting the Bruch membrane opening was designed.In the feature extraction stage,a residual structure with selective channel enhancement for feature maps is proposed based on ResNet.Adding a mechanism for scoring each channel of the feature map solved the problem of unbalanced attention of convolution neural network when the foreground is small in some extant.The multi-scale module is added in parallel to the backbone of the feature extraction network to solve the problem that the general convolutional neural network does not have diversity and comprehensiveness.Referring to the idea of the RPN by Faster-RCNN,the proposal of the optic nerve head region is added before the coordinate regression of the Bruch membrane opening,it solved the problem of extreme imbalance between the foreground and the background.The network is initialized by ResNet model trained on ImageNet using transfer learning and cross-validation is used to obtain test results.2.An improved U-Net structure UJ-Net is proposed to achieving the segmentation result of the retina and extract the inner limiting membrane.First,estimate the receptive field of the retinal region segmentation task and adjust the depth of the network.Secondly,an improved channel selection enhancement residual block is added to the encoder and decoder of the U-Net to solve the problem of the unbalance attention to the retina region.Then,considering noise in optic nerve head region and the low quality of image caused by poor optical signal intensity,a J-Net is added to eliminate the influence of noise by adding noise to the high dimensional feature map from the U-Net encoder.The accuracy of the segmentation result is improved by using multi-task learning.According to the detection of Bruch membrane opening and segmentation of inner limiting membrane,the position of the cup and disk boundary in single slice from the OCT image was determined,integrating all the slices containing Bruch membrane opening,and the cup-to-disk ratio is calculated by using the least square ellipse to fitting the cup and disk region.The test data set contained 22 well labeled primary Open-Angle Glaucoma(POAG)data.The Dice coefficient is 96.24%,Intersection over Union(IoU)is 95.67%,error is 0.058 and standard deviation is 0.021.Compared with the Cup-to-Disk Ratio(CDR)based on fundus digital images,our experimental result shows its reference value for clinical diagnosis of glaucoma.