Research On The Automatic Classification Of Diabetic Retinopathy

Diabetic retinopathy(DR)is one of the severe complication of diabetes and the main reason causing vision loss and blindness.The regular DR screening on the fundus images for early diagnosis and treatment has become an effective way to control the occurrence and the deterioration of the disease.Currently,DR are normally diagnosed and graded by ophthalmologists with manually examining and analyzing abnormal lesions on the fundus image(such as microaneurysms,hemorrhages and hard exudates),which is a time-consuming and laborious intensive process and restricts the large-scale DR screening.In addition,large screening populations and scarce ophthalmologists have become a bottleneck to meet the large-scale screening requirement.Therefore,the automatic DR classification method based on computer technology using the fundus images can quickly and effectively identify the DR lesions and classify the fundus image with the different severity of DR.It can not only free the doctor from the heavy manual reading work,but also effectively improve the accuracy,objectivity and rapidity of DR screening.And it has important clinical and social benefits for reducing the visual impairment of patients and promoting the implementation of large-scale DR screening in China.Many scholars and experts have conducted a lot of research on the automatic DR classification method at present.Because the DR classification method based on lesion detection has good clinical interpretation,the current research is mostly focused on the classification of DR based on related lesion detection.However,the research of these methods rarely pays attention to the impact of unbalanced data classification,incomplete annotation information of dataset and the inconsistent distribution of multi-source data,which makes the traditional supervised classification method difficult to obtain high accuracy.To this end,this dissertation starts from these problems and focuses on the realization of automatic DR classification based on fundus images.And it is studied from four aspects:the detection of microaneurysm,the diagnosis of DR,the multi-classification of DR and the multi-classification of diabetic macular edema.The main work of this dissertation are as follows:(1)In order to solve the issue of unbalanced data classification in microaneurysm detection,an adaptive over-sampling based ensemble classification method for microaneurysm detection is proposed.The approach firstly uses the mathematical morphology and region growing to detect suspected microaneurysms.In order to reduce the classification bias introduced by the unbalanced data in the classification of suspected lesions,this research proposes an adaptive over-sampling algorithm,which can adaptively determine the number of artificially synthesized samples of each minority sample.Furthermore,the sampling algorithm is combined with the ensemble frameworks of boosting,bagging,random subspace,and three adaptive over-sampling based ensemble classification models are constructed to realize the classification of suspected lesions.The publically available E-ophtha database are used to comprehensively test the proposed algorithm under the comparison with the state-of-the-art methods.The experimental results show the advantages and effectiveness of the proposed method in solving unbalanced data classification problems.(2)In order to solve the issue of the lesion labeling lack of dataset in the DR diagnosis,a DR diagnosis method based on multi-kernel and multi-instance learning is proposed.This method introduces multi-instance learning algorithm into the diagnosis of DR.The detected suspicious red lesion areas are considered as instances,and the whole image is considered as a bag,then a multi-instance learning model based on the kernel graph structure is adopted and integrated into the multi-kernel learning framework,Multi-instance model based on multi-core diagrams.Finally,a multi-instance learning model with multi-kernel graph is constructed to diagnosis of DR/no DR.In order to improve the classification performance and efficiency,the method also uses an optimization strategy of filtering the irrelevant instances to reduce the influence of irrelevant instances on multi-instance learning.The evaluation of proposed method is performed on publically available MESSIDOR database.Experimental results demonstrate that the proposed method can be used to diagnose the diabetic retinopathy efficiently without label information of the suspicious lesions,so as to avoid the time-consuming effort of labeling the lesion by specialists and false positive reduction.(3)In order to solve the issue of unbalanced data classification and inconsistent distribution of multi-source data in DR classification,a cost-senstive based semi-supervised ensemble algorithm for DR classification is proposed.Firstly,the suspicious red lesions containing microaneurysms and hemorrhages are detected.Secondly,in order to solve the problem of no lesion labeling of the target dataset in the classification of suspected lesions,this research uses semi-supervised learning technology to predict the categories of unlabeled samples based on the consistency of K-nearest neighbor samples and the strategy of high-confidence sampling.A combination of the semi-supervised learning method and a cost-sensitive based SVM in the ensemble framework of bagging is implemented,and a cost-senstive based semi-supervised bagging model is constructed to classify the microaneurysms and hemorrhages.Finally,the fundus images are graded into four stages of the DR severity based on the number of the different lesions.The publicly available MESSIDOR database are used to comprehensively test the proposed algorithm under the comparison with the state-of-the-art methods.Experimental results showed that the method proposed has substantial advantages over other methods.(4)In order to solve the difficulties in detecting optic disc(OD),macular and hard exudates(HEs)in diabetic macular edema(DME)grading,a DME classification method based on macular and hard exudate is proposed.The method firstly uses the multi-featured OD localization algorithm to detect the OD,and a region-based active contour model(RSF)based OD segmentation algorithm is used to obtain a fine OD contour.Then,a macular detection algorithm based on template matching is designed to achieve the center of the macula.Secondly,a method combined with vector quantization and the strategy of a local region-based segmentation is adopted to segment the suspected HEs,and then then an adaptive over-sampling based semi-supervised bagging classification algorithm is used to classify the suspected HEs.Finally,the fundus images are divided into three levels of DME based on the spatial distance between the HEs and the center of the macula.The evaluation of DME on publicly available MESSIDOR database is performed and the comparison with the state-of-the-art methods also is done.The experimental results show that the proposed method has a better accuracy and robustness.