A Study On Diabetic Retinopathy Detection Based On Supervised Classification

Diabetic Retinopathy(DR) is the serious ocular complications of Diabetes, which has become one of the main causes of the visually impaired and blind among adults. Diabetic Retinopathy can happen years without obvious effects on the vision. So the regular DR screening, timely diagnosis and treatment are very important. It can largely reduce the risk of vision loss. Fundus images obtained by fundus camera can accurately and objectively record fundus. Automated detection of Diabetic Retinopathy lesions(such as Hard Exudates, Microaneurysms and Haemorrhages) in fundus images can provide effective computer-aided diagnosis and lighten the burden of ophthalmologists.There are different anatomic structures(such as the blood vessels and optic disk) and lesions in Fundus image. Their similarity disturbs the detection of each other. In addition, the fundus imaging hardware condition also influences the quality of the image, increasing the difficulty of the detection. To solve those problems, we firstly conduct the fundus image preprocessing, and then segment normal the blood vessels and optic disk, finally we focus on the detection of lesions based on supervised classification method. In this paper, the main works are as follows:(1) Fundus image preprocessing. We preprocessing the fundus images by contrast enhancement and color specification technique, making the fundus images more suitable for subsequent detection of interested target.(2) The segement of blood vessels and optic disk. For the segmentation of blood vessels, we use mathematical morphology methods based on multi-orientations linear structure element to fastly segment the blood vessels, making use of the fact that the blood vessels are in linear form. For the segmentation of optic disk, firstly we narrowing the scope for the optic disk detection, making use of the relationship between the vascular distribution and the position of optic disk. And then we segment the optic disk by Hough transform circle detection methods as the optic disk presents as circular.(3) The detection of Hard Exudates(HEs). Combining the theories of feature extraction, morphology and Support Vector Machine, we propose a Hard Exudates detection method based on supervised classification. Firstly we coarsely segment the candidate regions by morphology and threshold methods and mask out the optic disk. Then we extract the color, edge, AM-FM texture and other features on the candidate regions to classify the HEs and non-HEs in the candidate regions by the Support Vector Machine classifier. Our method can eliminate the Soft Exudates and backgrounds which are similar to HEs in gray level, resulted in a better sensitivity and specificity.(4) The detection of Microaneurysms and Haemorrhages. As to the problems that Microaneurysms and Haemorrhages has low contrast and bad edges in fundus images, we propose a detection method based on Chan-Vese models and SVM. First of all we segment the candidate lesions making use of the local gray-scale differences after the fundus image preprocessing, and masked out the blood vessels from the candidate regions. Secondly we extract edge of the the candidate lesions by Chan-Vese model to have a better describe of the shape characteristic of lesions. Finally we extract color and shape features and eliminate the residual slender blood vessels and backgrounds by SVM classifer, achieving a precise detection of Microaneurysms and Haemorrhages.