Intelligent Diagnosis Of Glaucoma Based On Deep Learning

Glaucoma is a chronic and irreversible fundus disease,it is also the second leading cause of blindness worldwide.The pathogenesis of glaucoma is a chronic and unde-tectable process.Therefore,it is also called "silent vision killer".It is precisely because of the characteristics and great harm of glaucoma,the study of intelligent diagnosis of glaucoma is of great significance for the early prevention,detection and treatment of glaucoma,and it provides an important technical guarantee for people's quality of life.The intelligent diagnosis of glaucoma is mainly based on the analysis of digital fundus image,and the existing diagnosis methods mainly include two categories:the method based on optic cup and optic disc segmentation,and the direct classification method based on extracted features.The former is to segment the key areas(optic cup and optic disc)in the fundus images,and finally calculate the clinical parameters such as the ratio of vertical optic cup-to-disc,so as to provide the medical explanation for the subsequent classification diagnosis.In the latter,feature extraction is carried out for fundus images,and then the extracted features are used to train the classifier for direct classification diagnosis,avoiding the complex segmentation task.This thesis mainly studies the diagnosis methods of glaucoma based on these two types.For the former,this thesis mainly considers the important sub-problem of key areas segmentation.In recent years,with the continuous development of deep learning technology,the deep models with strong ability of automatic extraction of distinguishing features have achieved good performance in the diagnosis of glaucoma.The existing deep learning based key areas segmentation methods have some problems such as poor adaptability to geometric transformations in object and inadequate fusion of features at different levels.The existing methods based on feature extraction and direct classification have problems such as lack of data,class imbalance and domain shift.Besides,the data sets used for the diagnosis of glaucoma are small,and training data is also scarce.In order to overcome the above challenges,improve the performance of segmentation and classification,and meet the needs of rapid and accurate diagnosis in practical large-scale glaucoma diagnosis scenarios,this thesis proposes two methods of optic cup and optic disc segmentation methods based on modified U-net and one glaucoma classification diagnosis method based on deep convolutional neural network.(1)The core problem of diagnosis method of glaucoma based on optic cup and optic disc segmentation is how to segment optic cup and optic disc quickly and accu-rately.The existing segmentation algorithm based on deep segmentation network is dependent on the training of a large number of labeled data,because the model itself lacks the ability to adapt to geometric transformations in object.Due to the lack of glau-coma segmentation training data sets,this thesis introduces the deformable convolution operation to enhance the model's adaptability to geometric transformations in object from the inside,so as to improve the model segmentation performance.By introducing U-net++architecture,the features of different levels of the segmentation model can be better fused.In view of the difference in the difficulty of the segmentation task of optic cup and optic disc,this thesis adopts the architecture of different levels to ensure the segmentation performance and make it more efficient and lightweight.Experimental re-sults show that the two segmentation methods proposed in this thesis:DeforU-net and DeforU-net++-can effectively improve the segmentation performance of the network and improve the segmentation accuracy of the optic cup and optic disc.(2)The core problem of the diagnosis method of direct classification of glaucoma based on extracted features is how to extract discriminative features and use these fea-tures to train the classifier.The existing methods based on classification diagnosis have the problems of poor robustness and low classification accuracy,which are mainly due to the domain shift between different data sets in the actual diagnosis and the small size of data sets used for the diagnosis of glaucoma.In this thesis,the input images are pre-processed by U-net to extract the region of interest of the images,and then the image differences between different data sets are minimized through image enhancement to alleviate the problem of domain shift.By finetuning the pre-training network,the ro-bustness of the model is further improved.Experimental results show that this method can effectively improve the robustness and classification accuracy of the model.