Study On The Automated Detection Of Retinal Diseases Based On Deep Neural Networks

Vision is the most important way for people to get visual information from the world.With the growing popularity of smartphones and various electronic display devices,more and more people overuse their eyes.This results in a dramatic increase in myopia,which could further trigger myopic macular degeneration.On the other hand,with the increase of elderly parturient women,the number of premature infants has increased significantly,which leads to a high chance of retinopathy of prematurity(ROP).Screening of ROP is critical to preserve the vision of the infants.The health of eye has been identified by the World Health Organization(WHO)as one of the three major problems that affect the life quality of human being.As a newly developed machine learning method,deep learning has achieved many successful applications in the field of retinopathy screening.However,there are still significant limitations of the existing methods when dealing with complex and diverse retinopathy.This prompted us to continue to research thoroughly the challenging issues in deep learning methods for the application in retinopathy screening.The main contributions of the thesis are the following:(1)We proposed a unified dropout method framework ?-dropout.The dropout method based on different sampling distribution is unified into a dropout framework based on ? distribution,and then the dropout method selection problem is changed to a parameter tuning problem.On this basis,as an example,an adaptive ?-dropout is proposed.Experimental results on several different datasets show that adaptive ??dropout can obtain better results on different datasets than the previous best methods,which verifies the robustness and advancement of the proposed method.(2)An improved U-shaped fundus vessel segmentation model is proposed.We use the whole fundus image as the input of the network to make full use of the tree-like structure characteristics and global location information of fundus vessels.We introduce a feature weighting mechanism to achieve efficient extraction and utilization of features.Meanwhile,this mechanism also avoids the influence of unimportant features in the training of vascular segmentation network.In order to further improve the image representation ability of the convolutional network,the weighted multi-level cross entropy loss is used as the loss function of the segmented network.Compared with typical methods on the public dataset,the method can solve the problems of the loss of fine blood vessels,the blur of vessel segmentation boundary and vessel fracture,and obtain better results for segmentation of blood vessels in fundus images.(3)For Plus disease in retinopathy of prematurity,an end-to-end Plus disease diagnosis system is developed based on the improved U-net and Densenet architecture.The system firstly extracts the blood vessels from fundus images of prematurity using the blood vessel segmentation model and then feeds the segmented blood vessel images into Densenet network to diagnose Plus disease in retinopathy of prematurity.In addition,a quantitative analysis method of Plus disease is proposed.A method for quantitative analysis of Plus disease in retinopathy of prematurity was proposed to quantitatively analyze the condition of Plus disease by evaluating the vascular curvature,vascular width,and fractal dimension.The experimental results proved the validity and accuracy of the proposed Plus disease diagnosis system and Plus disease quantitative analysis method.(4)An assisting diagnosing system for detecting myopic macular degeneration and analysis the severity of myopic macular degeneration were proposed.The system consists of two levels of network.The first level of network is used to determine whether there is myopic macular lesion.The second level network is used to identify the severity of myopic macular lesions.Experimental results show that the proposed method can automatically extract effective features to detect myopic macular degeneration and identify the severity of the disease.The specificity and sensitivity of the proposed method are over 0.9.