Abstract
Purpose :
To use an AI-based fluid algorithm to determine the discriminative ability of eye-care professionals in grading intra- and subretinal fluid (IRF, SRF) presence and volume change in real-world OCT images of patients with nAMD.
Methods :
Patients with nAMD treated at our retina department between 2007 and 2018 were included in the Vienna Imaging Biomarker Eye Study (VIBES) registry. 5 retinologists (RET), 3 ophthalmology residents (RES), 3 general ophthalmologists (GENO), 3 orthoptists (ORTH) and 3 certified graders from the Vienna Reading Center (VRC) graded the presence of IRF/SRF at two consecutive visits as well as the change (increase/no change/decrease) on Spectralis OCT images. Intervals ranged from 28 to 120 days between visits and anti-VEGF injections were given between the two visits. The Vienna Fluid Monitor Version 2 (RetInSight, Vienna, Austria) was applied to automatically segment and quantify IRF/SRF volumes. ROC curves were then generated to determine volume cut-offs (nl) for fluid presence and change that provided the best tradeoff between sensitivity (Sen) and specificity (Spe) (=Youden Index).
Results :
124 visit pairs of 59 eyes were included. For IRF presence, fluid volumes between 3,3-6nl were most accurately detected across all groups with 0,79-0,87 (Sen) and 0,92-0,97 (Spe). For SRF presence, volumes between 4-7,1nl correlated with 0,82-0,95 (Sen) and 0,92-0,97 (Spe). In terms of fluid change, an IRF increase as small as 1,7-3nl was detected with 0,91-1,0 (Sen) and 0,88-0,93 (Spe) and an IRF decrease of 2,8-16,9nl with 0,83-0,9 (Sen) and 0,9-0,96 (Spe). SRF increases of 0,6-6,4nl were detected with 0,84-0,94 (Sen) and 0,89-0,96 (Spe), SRF decreases of 6,4-9,3nl with 0,93-0,98 (Sen) and 0,87-0,95 (Spe). The AUC was high for both fluid presence and fluid change with values consistently over 0,89 across all groups.
Conclusions :
AI-based fluid analysis permits to determine the accuracy of human experts from different professional backgrounds in grading retinal fluid types and their changes over time. The introduction of such objective, quantitative parameters results in a consistent and exact management, irrespective of the treating physicians’ background.
This abstract was presented at the 2023 ARVO Annual Meeting, held in New Orleans, LA, April 23-27, 2023.