B017: A comparison of manual and 3 automated methods of determining retinal arteriolar diameters

• Published in: American journal of hypertension 2000-04, Vol.13 (S2), p.192A-193A
• Main Authors: Chapman, J.N., Witt, N., Stanton, A.V., Gau, X., Thom, S.A., Hughes, A.D.
• Format: Article
• Language: English
• Subjects: automated image analysis; retinal arterioles; Vascular network geometry
• ISSN: 0895-7061; 1941-7225; 1879-1905
• DOI: 10.1016/S0895-7061(00)00678-6

The retinal circulation is easily visible and affected by cardiovascular disease including hypertension and diabetes. We have previously reported that measurement of retinal arteriolar geometry (defined by bifurcation angles [ω] and junction exponents [×]) allows early assessment of hypertensive changes in the retina. However, manual measurements are time consuming and subjective and computer-based quantification should offer improved speed and reproducibility. We therefore compared the accuracy and reproducibility of 3 computer algorithms designed to measure arteriolar diameters from red-free retinal photographic negatives. These results were compared with diameters measured manually from red-free and fluorescein angiograms of the same subjects. The algorithms evaluated were: 1) fitting vessel intensity profiles to a double Gaussian function by non-linear regression, 2) a standard edge detection algorithm (‘Sobel’) and 3) determination of points of maximum intensity variation by a sliding linear regression filter (‘SLRF’). Using each technique 60 arteries were measured at 5 adjacent sites and the median diameter (pixels) calculated. Data were analysed by Bland-Altman plots and are presented as mean [95% CI] differences compared with reference. Mean vessel diameter measured from fluorescein angiograms was 21.17. Mean differences in diameter (fluorescein—red-free test diameter) were: manual = 1.95 [0.91–3.00], Gaussian = 2.73 [−0.13–3.31], Sobel = 4.84 [2.62–7.07] and SLRF = −1.06 [−2.34–0.24]. To compare internal reproducibility a further 21 vessels were measured on 2 occasions by each method. Differences between measurements were: manual 0.60 [0.06–1.14], Gaussian −0.12 [−1.59–1.34], Sobel 1.35 [0.44–2.25] and SLRF −0.12 [−1.06–0.83]. In conclusion, of the three automated methods, the SLRF method was the most closely in agreement with reference methods, the most reliable and the most reproducible. The Sobel method was least reliable due to frequent detection of the light reflex as the vessel edge. Application of automated approaches to retinal arteriolar network analysis may prove useful in cardiovascular risk assessment.