| Diabetic retinopathy(DR)is a kind of serious eye complication of diabetes,which can lead to visual impairment or even irreversible blindness.Many studies have shown that periodical inspection of the fundus can early detect the lesions,and early diagnosis and treatment can effectively prevent vision loss or blindness.The fundus camera has the advantages of low cost,convenience and noninvasive detection,and is widely used in large scale diabetic retinopathy screening.Microaneurysms(MAs)are the first sign of diabetic retinopathy,which can be observed in retinal fundus image.Therefore,the automatic detection of MAs is of great significance to the early prevention and treatment of diabetic retinopathy by using computer image processing technology.Retinal fundus images are often with non-uniform illumination,poor contrast and noise image.In addition,the size of MAs is small and variable,and is also easy to be disturbed by the background image.Thus,the automatic detection of MAs is still a challenging problem in practice.In this thesis,we propose an automated method to detect MAs in retinal fundus images,and the main work and contributions are summarized as following:Firstly,image preprocessing of the fundus image is performed to produce a more suitable image for MAs detection,including extracting ROI(Region of Interest),illumination equalization,contrast limited adaptive histogram equalization(CLAHE),Gaussian filter and detection of optic disc.Secondly,MAs candidate extraction is necessary and important.We propose a new method for extracting candidates,which are similar to MAs,based on dynamic shape analysis.Through calculating roundness,area and depth information and setting determination condition,we can eliminate those unsatisfactory candidates.We test the method on ROC(Retinal Online Challenge)training dataset,and the results show that the method can effectively extract most real MAs.Thirdly,we analyze the MAs candidates from the profile and local angles and a total of 48 characteristic features which contain 8 profile features and 40 local features are extracted.Finally,an AdaBoost classifier is trained based on the above 48 features to distinguish true MAs from spurious candidates.The proposed method has been evaluated on ROC public dataset.The experimental results demonstrate the efficiency and effectiveness of the proposed method,and its potential to be used to in the clinical diagnose of DR. |
