Functional modifications of retinal images in quantitative plus disease diagnosis of retinopathy of prematurity

Abstract

Purpose: Plus disease is a key indicator for treatment-requiring ROP, but significant variability exists among ROP experts in plus disease diagnosis. Computer-based image analysis systems have emerged as an alternative paradigm that can quantify vascular features of retinal images to decrease subjectivity in plus disease diagnosis. The purpose of this study was to examine how various properties of retinal images influence the quantification of plus disease.

Methods: Wide-angle retinal images from infants with ROP were interpreted as either “plus”, “pre-plus”, or “not plus” based on a panel of 3 or more readers trained in ROP diagnosis. Retinal vessels from all images were manually segmented by author consensus. Each image was then cropped into circular and rectangular shapes based on the center of the optic disc. Circular crops originated on the center of the optic disc with an integer radius ranging from 1-6 disc diameters (DD). Rectangular crops were designed such that the center of the optic disc was fixed at 1/3 of the horizontal length of the image with the following dimensions: rectangle with a height of 3DD x width of 4DD (3ddHx4ddW), 4.5ddHx6ddW, 6.75ddHx9ddW, 9ddHx12ddW, and 12ddHx15ddW. 5 wide-angle retinal images were selected (3 pre-plus, 2 plus), and each was cropped 11 times, for a total of 60 images for analysis. Different vessel-level tortuosity measures, calculated using an image analysis system designed by the authors, were combined with a supervised approach to provide a global image-level tortuosity index.

Results: There was a strong positive correlation between the sizes of rectangular crops and the degree of tortuosity as a function of distance from the optic disc center (r_s(25) = 0.643, p < 0.001). Tortuosity values based off of rectangular shapes were a significant predictor of the clinical diagnosis of pre-plus or plus disease (OR: 1.50, 95% CI:1.04 - 2.17; p = 0.032). There was no statistically significant relationship between the clinical diagnosis and tortuosity values based off of circular shapes (OR: 1.12, 95% CI: 0.94-1.34; p = 0.214).

Conclusions: Computer generated tortuosity measurements using rectangular crops that incorporate more temporal and nasal vessels may have greater prognostic value in diagnosing plus disease. This finding may have implications in the clinical diagnosis of plus disease and in the design of computer-based image analysis systems.

Keywords: 706 retinopathy of prematurity • 550 imaging/image analysis: clinical