An Automatic Method For The Artery/Vein Classification In Retinal Images

Retinal vessel is the only part of microcirculation system that can be observed directly by non-invasive means. Retinal vascular structure can be observed clearly in retinal images.Morphological changes of retinal vascular in retinal image such as arteriolar-to-venular diameter ratio(AVR) and arteriovenous(AV)-nicking are signs of many systemic diseases,for example, diabetes, hypertension, coronary heart disease and heart disease.Retinal vascular arteriovenous(AV) classification and arteriovenous nicking detection in color fundus images are popular research topics.The accuracy of AV-classification varies in retinal images with different resolutions. Manual identification of AV-nicking is very complicated, which requires well-trained graders. It is also time-consuming and subjective.Algorithms to classify artery/vein and detect AV-nicking are proposed in this thesis.The algorithm of AV-classification is based on the feature of vascular color,morphology and topology. A four piecewise Gaussian model is proposed to profile the cross section of a vessel and the model parameters are employed as the features for classification.A vascular segment is classified as artery or vein using SVM. Then, a voting scheme is applied on the classification results of vessel segments to decide the vessel type of a whole vessel. In the algorithm of AV-nicking detection, vessel segmentation, centerline extraction and intersection point detection are performed. A method of vessel diameters measurment is proposed after the voids, hidden vessels and micro-vessels in the segmentation are removed, which can describe the continuous change of the vein caliber from the cross region. A plenty of detection features can be extracted based on the vein diameters. The AV-nicking crossing point can be identified, according to a rule-based criteria.The databases of DRIVE and Wisconsin were used to test the proposed algorithm of AV-classification. The algorithm of AV-nicking detection was tested by the clinical images.The result shows that our algorithm can classify the artery/vein and detect the nicking points successfully in most cases.