Fundus Image Analysis Based On Machine Learning

The fundus tissues such as optic cup,optic disc and central retinal artery and vein observed in fundus images provide an important basis for the early screening and diagnosis of ophthalmic diseases,diabetes,cardiovascular and cerebrovascular diseases.These diseases usually have no obvious symptoms in the early stage,but the patient's fundus tissue may show different degrees of lesions.For example,glaucoma usually leads to optic disc atrophy and depression? The early clinical manifestation of hypertensive retinopathy is retinal arteriosclerosis? The early clinical manifestations of diabetic retinopathy include retinal hemorrhage and microaneurysm formation.It is not difficult to find that most eye lesions are closely related to the structure of fundus tissue.In ophthalmic diagnosis and treatment,clinical ophthalmologists can evaluate and diagnose the eye health status of patients through fundus images.However,in the face of the evaluation and diagnosis needs of a large number of fundus images,there is no doubt that manual detection has some disadvantages,such as slow speed,poor repeatability,high labor intensity and so on.Therefore,the research and development of intelligent and automatic fundus image processing methods and quantitative analysis technology is of great clinical significance for early screening,treatment monitoring and postoperative estimation of ophthalmic diseases and related diseases.Under this background,with the premise of contributing to clinical diagnosis and treatment and the goal of realizing computer-aided diagnosis,this dissertation makes an in-depth study on the joint segmentation method of optic cup and optic disc,retinal vessels segmentation method and hierarchical division method of retinal vascular network in fundus image from two aspects of theory and application.The main contributions and innovations of this dissertation are as follows:1.A cup-disk joint segmentation network model(CDJS-Net)is proposed.The model combines the convolutional neural networks with the generative adversarial networks to jointly segment the optic cup and optic disc in the fundus image.Through the discrimination,feedback,optimization and adjustment of relevant parameters,the model achieves autonomous learning and optimization.Experiments and analysis conducted on public dataset demonstrate that the segmentation performance of CDJS-Net on optic disk is superior to the existing segmentation models.2.A multi-scale iterative aggregation u-network model(MIA-UNet)is proposed.For the segmentation of retinal vessels,the model redefines and designs the interconnection between encoder and decoder,as well as the internal connection between decoder subnetworks in classic U-Net,which makes the aggregation of full-scale semantic features possible.Experiments and analysis conducted on three public datasets demonstrate that the MIA-UNet is better than the existing retinal vessel segmentation models.3.A hierarchical division framework of retinal vascular network(HDF-RVN)is proposed.The framework consists of three stages: in the first stage,the retinal vessels and optic disc in the fundus image are segmented? In the second stage,the vascular tree is extracted and generated from the segmented retinal vessels? In the third stage,the hierarchical features of vascular bifurcation points are classified,so as to realize the hierarchical division of retinal vascular network.Extensive experiments and analysis conducted on two public datasets demonstrate the enforceability of the proposed hierarchical division framework.