Abstract
Purpose :
Software for automated vascular parameterisation is capable to quickly analyse large amounts of images.The purpose of this project was to assess agreement between automated parameterisation and manual vascular classification to identity error sources
Methods :
Retinal images from 626 diabetic patients with varying degrees of diabetic retinal disease (Institute for Research in Ophthalmology,Poznan, Poland) and a random sample of 314 individuals from ocular healthy participants in the Leipzig Research Centre for Civilization Diseases-LIFE Adult study, Leipzig, Germany, were analysed using a semi-automatic vessel analysis programme (MonaReva,Belgium). Central retinal artery equivalent (CRAE), central retinal vein equivalent (CRVE), arterio-venous ratio (AVR) and vessel density were calculated using the standardized protocol in a concentric measurement zone around the optic disc. Following automated analysis, all images were manually checked to assess if arterioles and venules have been classified correctly and accurately. Missing, misclassified/segmented vessels were manually reclassified/segmented and added/deleted vessels accordingly. Measurements from automated and manually checked data were compared using Bland-Altman-Analyses and regression analyses to identify proportional bias
Results :
The source of misclassification of calibers and type is similar in both cohorts and often due to arteries and veins being located very close to each other. Differences in vessel density measurements between automated and manually checked segmentation occurred due to inner limiting membrane reflection in healthy individuals, whereas in diabetics was mainly due to neovascularisations and previous laser treatment lesions in the retina. Bias and agreement (between automated and manually checked parameterisation) was poor for CRAE, CRVE and AVR in both groups (Fig1). Bias and agreement for vessel density was poor in diabetic patients but negligible in healthy individuals
Conclusions :
Inaccuracies of vessel parameterisation,i.e. arteries classified as veins,two vessels identified as one,nerve fibre reflection and choroidal vessel segmentation can impact decision making for screening purposes and when computing cardiovascular risk scores. For screening purposes, it is important to identify markers which provide a high sensitivity and specificity to correctly segregate between individuals with different risk
This abstract was presented at the 2024 ARVO Annual Meeting, held in Seattle, WA, May 5-9, 2024.