Abstract
Purpose: :
To automatically determine the arteriolar-to-venular width ratio (AVR) of the retina. Retinal vessel characteristics, including central retinal arteriolar equivalent (CRAE), central retinal venular equivalent (CRVE), and AVR are currently being used in large-scale studies to evaluate the relationship to other systemic pathologies such as cerebral lacunar infarcts, cognitive decline, coronary artery disease and depression. However, each of these is based on a semi-automatic technique that relies on a human grader to edit the measurements as needed. We tested an automated system to calculate AVR and compared it to a widely-used semi-automated method.
Methods: :
A set of 20 de-identified, high resolution digital color fundus photographs centered on the optic disc were acquired from 20 subjects with glaucoma at the University of Iowa Hospitals and Clinics using a sequential variable-base stereo fundus camera (Carl Zeiss Meditec, Dublin, CA). To automatically determine the AVR, a computer system was developed that performed the following operations: optic disc location detection, vessel segmentation, accurate vessel width measurement, vessel network analysis and artery-vein classification. Optic disc detection was necessary to locate the region of interest (ROI) in which the measurements were obtained. After finding the ROI, the system started the automatic vessel segmentation. A separate algorithm precisely measured the local vessel widths. Finally, the system determined the type (artery or vein) of the vessels found within the ROI using a classification approach. By applying the formulas previously described [Knudtson et al. (2003), Curr Eye Res. 27(3):143-0] the system determined the AVR using the 6 widest arteries and 6 widest veins.
Results: :
The results of the automatic method were compared with those of the semi-automatic computer method, IVAN (The University of Wisconsin, Madison, WI). The mean AVR of the automated system was 0.68 (SD 0.08) while the mean AVR of the semi-automated system was 0.67 (SD 0.09). The overall mean absolute error was 0.06 with a standard deviation of 0.04.
Conclusions: :
Automated determination of the AVR using a computer system is possible with a relatively small error, yielding comparable results to the semi-automated system currently in use. This method may provide an additional tool in the investigation of retinal vascular diameter changes as a marker for systemic disease, however validation of this system in a larger dataset is still needed.
Keywords: image processing • clinical (human) or epidemiologic studies: systems/equipment/techniques • retina