Research On Glaucoma Detection Method Based On Deep Convolutional Neural Network

Glaucoma is one of the common blinding eye diseases,and its damage to vision is irreversible.If patients cannot be diagnosed and treated in time,there is a risk of blindness.The accuracy of glaucoma detection indicators directly affects the results of diagnosis.Traditional methods mainly rely on the diagnostic experience of professional ophthalmologists,which limits the development of glaucoma detection.The development of automated methods for glaucoma detection has gradually become the mainstream trend.At present,the automatic detection methods of glaucoma are mainly divided into two categories,one is a classification method based on image features,and the other is a segmentation method that calculates the cup-to-disc ratio by segmenting the optic cup and optic disc area in the image.The two methods solve the problem of glaucoma detection from different angles.With the rapid development of artificial intelligence technology and the maturity of deep neural networks,more and more researchers are turning their attention to the application of deep neural networks in medical images.The focus of this thesis is to explore methods for automatic glaucoma detection using deep convolutional neural networks.Image classification methods based on deep convolutional neural networks have performed well in various tasks.In this thesis,deep convolutional neural networks are used for glaucoma classification tasks.The core problem of the glaucoma detection method based on feature extraction and direct classification is whether the network can extract the most discriminative glaucoma classification features.The most discriminative feature of glaucoma classification is contained in the key area of the optic nerve head,which accounts for a very small proportion of the entire fundus map,about 3%,and this area is also affected by interference information such as blood vessels and optic nerves.In view of this feature of glaucoma classification,the network needs to quickly and accurately locate key regions,strengthen the information in the region,suppress other irrelevant information,and extract the most discriminative features of glaucoma classification.Based on the mature image classification network Res Net50,this thesis improves the structure of its network and introduces the attention mechanism to make it locate the key areas quickly and accurately.By introducing grouped convolution building blocks,the feature extraction capability of the network is enhanced without increasing the network parameters.The experimental results show that the classification performance of the method has been significantly improved,and its accuracy,sensitivity and specificity have reached 95.2%,95.6% and 94.3%,respectively,and the training efficiency has been significantly improved.The cup-to-disk ratio is an important indicator in the process of glaucoma detection.To obtain an accurate cup-to-disk ratio,it is necessary to accurately segment the optic cup and optic disc,which is also the key to the glaucoma segmentation method.In clinical practice,the optic cup and optic disc are segmented manually to calculate the cup-to-disk ratio,and its accuracy and efficiency are difficult to guarantee.Therefore,it is urgent to develop an efficient and accurate automatic optic cup and optic disc segmentation system.The existing methods have the problems of insufficient fusion of different levels of features and low accuracy of optic cup segmentation.In order to solve the above problems,this thesis proposes a deep neural network named R2RADF-net,which can jointly segment the optic cup and optic disc region in the fundus image by fully mining the complementary information encoded at different levels.In addition,the introduction of multi-scale input structure enables the network to obtain receptive fields of different sizes,and the recursive residual convolution module is combined to improve the performance of joint segmentation.Extensive experimental results on the public REFUGE dataset show that R2RADF-net has achieved significant performance improvement over existing methods,with the cup overlap error and cup-optic disc balance accuracy of 0.088,0.967,and 0.960,respectively.