Abstract
Abstract: :
Purpose: To develop a method for quantitative assessment of the retinal vessel system. Method: The assessment is based on arterio-venous ratio (AVR) calculated as a ratio of average arteriolar diameter and average venous diameter. The method consists of four steps: measurement zone assignment, vessel recognition, vessel classification and arterio-venous ratio calculation. Images used in the study are images of non-mydriatic retinal camera, 760x570 pixels, 24 bits per pixel (standard RGB), field of view 22,5° and 45°. The algorithm works with the green channel of the images except for vessel classification where the red channel is used. The measurement zone is defined between 1 and 2 optic disc radius from optic disc margin. The assignment of this zone is based on outlining of the optic disc by a circle with the means of the Hough transform. Then the green channel is corrected for nonuniform illumination. The vessel recognition is based on isocontour calculation (analogous to the geography) and tracking algorithm (vessel are tracked and at bifurcations and crossings decisions are made how to continue). The vessel classification into arteries and veins is based on information in the red channel. The channel is corrected for nonuniform illumination and isocontours are calculated. The isocontours of veins are deeper, which reflects the fact that veins are darker than arteries. Since we are able to recognize bifurcation and crossing, this knowledge serves as support criterion for classification. AVR is calculated as ratio of average arteriolar diameter and average venous diameter of the vessels within measurement zones. Diameters were not calculated at bifurcations or crossings and for vessels with a poor quality (automatically defined by the algorithm). Vessels with small diameter were excluded, because it is not even possible for an ophthalmologist to distinguish between arteries and veins. Results: The developed system calculates AVR correctly if the quality of the red channel is good. In such cases vessel classification (arteries/veins) is 100% correct. 15 images were used for training (development) and 248 for testing. Conclusions: The first prototype of the system for automated AVR is available. The most difficult task of the algorithm, vessel classification into arteries and veins, can be solved if the red channel is of a good quality. If the red channel is of a poor quality, the additional criterions such crossings or vessel color have to be applied (it is currently investigated).
Keywords: imaging/image analysis: non-clinical • clinical (human) or epidemiologic studies: ris • retina