Quantifying vascular tortuosity in retinopathy of prematurity: impact of segmentation method and vascular length

Abstract

Purpose: Plus disease is defined by vascular tortuosity and dilation, and is a critical marker for treatment-requiring retinopathy of prematurity (ROP). Various computational systems for quantifying retinal vascular tortuosity have been developed, which rely on different algorithms for segmenting vessels and calculating tortuosity. The variability in tortuosity values calculated by these systems is not known. This is important to understand the extent to which quantitative vascular tortuosity in ROP is dependent on the specific computer system used. The purpose of this study is to compare the mean tortuosity calculated by two different systems, which used different methods of vascular segmentation but similar algorithms for tortuosity.

Methods: 34 wide-angle retinal images from infants with ROP were analyzed using two different systems: (1) System A performed semi-automated segmentation and calculated tortuosity over the entire vessel length, and (2) System B relied on manual segmentation and calculated tortuosity as a function of distance from the optic disc center. Both systems calculated tortuosity using the same algorithm: vessel length from start point to end point divided by the straight line distance between the same points. Overall tortuosity of each image was used for comparison, and was defined as mean tortuosity of all individual arteries and veins segmented by each system. Correlation between the two systems was measured by Pearson’s r coefficient, and correlation coefficients were calculated for varying distances from the optic disc.

Results: Systems A and B showed strong correlation when computing tortuosity over the entire vessel length. System A (analyzing the entire vessel length) had strong correlation with System B (analyzing over a range of vessel lengths): for 1 disc diameter (DD) distance (r=0.688, p <0.0001), 2 DD (r=0.801, p <0.0001), 3 DD (r=0.811, p <0.0001), and 4 DD (r=0.819, p <0.0001) from the optic disc margin. There was not a statistical difference between the correlations calculated (p >0.22).

Conclusions: There is strong correlation between vascular tortuosity computations in ROP when performed using similar methods by different systems, even when using a range of vascular lengths and segmentation methods. This may have implications for development and implementation of computer-based image analysis systems for ROP.

Keywords: 706 retinopathy of prematurity