Purpose:
Analysis of morphologic changes prominent either to a retinalartery or a vein may be a diagnostic indicator of a retinopathyor a systemic disease. Hence, it is desirable to identify theretinal vessel trees and classify them as belonging to an arteryclass or a vein class.
Methods:
We apply the artery-venous classification method to a datasetof 15 fundus images selected randomly from normal and diabeticretinopathy subjects. A method described by our group for thestructural mapping of retinal vessel trees, may form a baselinefor artery-venous classification system (SPIE’ 2011).We develop an algorithm that classifies the separated vesseltrees using color features and anatomic property of artery-venouscrossing. The feature vectors consisting of four features, viz.,mean and standard deviation of each of green channel and huechannel, from the 3x3 neighborhood of vessel centerline pixelsare acquired. Based on the feature vector and artery-venouscrossing property, the algorithm classifies the vessel treesinto an artery class or a vein class, using fuzzy C-means clusteringalgorithm.
Results:
Figure shows the classified vessel trees, labeled with red colorfor arteries and blue color for veins. Application of the proposedmethod over the dataset resulted into an accuracy of 88.28%correctly classified vessel pixels, when compared with the manuallylabeled ground truth.
Conclusions:
The artery venous classification results match well with theground truth, suggesting that the method has potential in vesselnetwork analysis.
Keywords: image processing • imaging/image analysis: non-clinical • retina