Automated vessel density detection in fluorescein angiography images correlates with vision in proliferative diabetic retinopathy

Abstract

Purpose : To use an artificial intelligence algorithm to automate vessel detection from input FA images and conduct a secondary analysis of data from a prospective randomized controlled trial to investigate the correlation between quantifiable vessel density and vision in patients with proliferative diabetic retinopathy (PDR).

Methods : Secondary analysis of a prospective randomized controlled trial (Intravitreal Aflibercept for Retinal Non-Perfusion in Proliferative Diabetic Retinopathy [RECOVERY]). We designed and trained an algorithm to detect retinal vessels in FA images in an automated fashion. We then used our algorithm to study the correlation between baseline vessel density and best corrected visual acuity (BCVA) and central retinal thickness (CRT) for study patients with PDR without significant center-involving diabetic macular edema (CRT ≤320µm). Reliability of the algorithm was tested using the intraclass correlation (ICC).

Results : 42 patients from the RECOVERY trial who had both baseline FA images and optical coherence tomography (OCT) data were included in our study. Mean time for automated vessel detection from a FA image is 22.1 seconds (standard deviation 0.21 seconds). Our algorithm analyzed FA images with a reliability measure (ICC) of 0.98. The framework for correlating vessel density with BCVA and CRT is depicted in Figure 1. A positive correlation (r = 0.4071, p = 0.0075) was found between baseline macular vessel density and BCVA in PDR patients. Example detected vessel maps from FA regions of two patients who have high BCVA and two patients who have low BCVA are shown in Figure 2. No correlation was found between vessel density and CRT, however, the study excluded patients with CRT > 320µm.

Conclusions : Our algorithm is capable of reliably quantifying vessel density in a fully automated fashion from baseline FA images. We found a positive correlation between computed vessel density and BCVA in PDR patients without center-involving macular edema, but not CRT.

This is a 2020 ARVO Annual Meeting abstract.