Longitudinal Precision of Vasculature Parameter Assessment on Ultra-widefield Fluorescein Angiography Using a Deep-Learning Model for Vascular Segmentation in Eyes without Vascular Pathology

Abstract

Purpose : Longitudinal comparative measurements in ultra-widefield angiography (UWFA) can be confounded by multiple variables, including amount of fluorescein injected, eye exposure, image contrast, and image timing. Repeatability of longitudinal vasculature analysis has not been extensively explored. The purpose of this study was to evaluate the longitudinal consistency of retinal vasculature measures in phase-matched UWFA images utilizing a deep-learning augmented vascular segmentation system in eyes without retinal disease.

Methods : Eyes were identified with consecutive UWFA images. Following identification of the optimal phase-matched pairs, projection correction image alignment was performed and a common region of interest (ROI) capturing the same vasculature was determined for each pair. A trained automated deep-learning algorithm extracted the retinal vasculature from UWFA pairs. Geodesic diameter (GD) maps were created using ImageJ MorphoLibJ to measure the average distance between one central point (located at the macula or optic disc) and any point within the vascular network. Vasculature masks and GD maps were analyzed for vascular area, length and GD parameters.

Results : Twelve eyes of nine patients without any retinal vascular disease were identified and phase-matched. Mean total vessel area, length, disc-centered and macular GD were respectively 75.9E+4 ± 16.7E+4, 14.6E+4 ± 4.0E4, 971 ± 39 and 1090 ± 47 pixels in first time-point and 74.0E+4 ± 15.3E+4, 14.2E+4 ± 3.4E4, 963 ± 35 and 1084 ± 49 pixels in second time-point. No significant differences between time-points were identified in measures of vessel area, length and GD parameters.

Conclusions : Phase-matched UWFA longitudinal vasculature masks in eyes without vascular pathology demonstrate consistent measurements over time utilizing a deep-learning augmented segmentation model. Additional validation of vascular UWFA measurement stability over time in normal eyes will be explored in a larger dataset in the future, but this technique appears to hold promise as an opportunity for vascular feature interrogation in UWFA.

This is a 2020 ARVO Annual Meeting abstract.

 

Phase matched UWFA images (A, B) and their corresponding vasculature masks (C, D), color scaled disc centered GD maps (E, F), and color scaled macula centered GD maps (G, H) of an eye without vascular pathology.