Application Of Deep Learning In Glaucoma Census System

With the massive accumulation of various types of data in modern society,the improvement of computer computing capabilities,and the improvement of algorithm models and frameworks in academia and industry,the development of deep learning is making rapid progress.At present,deep learning has achieved many results that have been applied in all walks of life and solved many complex problems.Due to the uneven distribution of medical resources in China,conventional fundus cameras are not convenient to conduct eye examinations for people in remote areas.Therefore,it is necessary to use portable handheld fundus cameras for identity authentication and eye disease detection with deep learning technology,it not only saves medical and health resources,improves the working efficiency of medical staff,but also relatively balances the uneven distribution of medical standards,the uneven level of ophthalmologists,and the difficulty of extending large medical equipment.Aiming at the background of eye disease census,this article mainly implements the identity verification process and glaucoma-assisted detection process in the census system.The specific research content is as follows:1.Use the prediction results of the positioning network to detect bad images and locate the center point of the optic disc.The positioning network uses the architecture design of the existing network Faster RCNN,the main changes include changing the feature extraction network,adding multi-scale input layers,and adjusting the selection of anchors for region proposal network.Because the optic disc area itself is relatively small compared to the entire fundus image,the modified network is more suitable for small object detection.2.Use the mathematical morphology method to enhance the blood vessels in the fundus image during the identity verification process,and use the improved FaceNet that add classification units during the training process to implement fundus image identity verification,so that features can be extracted more quickly and efficiently.Through experiments,the authentication method of this article can reach 97.5% accuracy on the test set.3.A preprocessing method is proposed for the fundus image.By analyzing the channel information,only the effective channels are retained,and the existing image processing methods are used to increase the clarity and resolution of the optic cup and optic disc.4.In this paper,an end-to-end deep network named M-FCN to segment the optic disc and optic cup automatically is designed.M-FCN adjusts the basic network structure on FCN,and increases the design of multi-scale models.The input changes from a single original fundus image to a multi-scale pyramid image.The experimental results prove that the designed M-FCN has strong feasibility and advantages.