Abstract
Purpose :
Age-related macular degeneration (AMD), diabetic retinopathy (DR), and glaucoma are vision-threatening diseases (VTDs) affecting 36 million people in the USA. With 5.7 ophthalmologists per 100,000 Americans, over 50% of VTDs go undetected. We assessed deep learning Artificial Intelligence (DLAI) in VTD detection in community and clinical settings.
Methods :
223 subjects (mean age 54.6, 58% male) from community screenings (A) and clinic (B) underwent 45-degree retinal imaging. In A (non-dilated), an onsite telemedicine reader (R1) and remote ophthalmologist (R2) graded image quality (gamma and alignment, 1-5 scale) and referable VTD using the international grading scales for AMD and DR, and cup-to-disc ratio and nerve fiber layer for glaucoma. In B (dilated), gradings were collected from R1 and the clinical diagnosis (d). A senior ophthalmologist (R3) adjudicated disputed findings. In A, DLAI VTD referral was compared to R1/R2/R3 consensus (S); in B, overall referral was compared to R1/d/R3 consensus (C). Images were uploaded to a cloud-based DLAI (SELENA+, EyRIS Pte Ltd) (Fig 1). Cohen’s kappa assessed intergrader agreement.
Results :
R1 and R2 found 4.7% eyes ungradable. DLAI marked 55.6% ungradable; 74.6% of them were for AMD. Of the DLAI ungradable eyes, image quality was ≤ 3, and 56.2% had ≥ 1+ cataract (R1). Compared to in A, in B DLAI had higher sensitivity (97.1% vs. 63.2%) and positive predictive value (69.4% vs. 32%). In A, DLAI had higher specificity (94.5% vs.16.7%) and negative predictive value (98.4% vs. 75.0%) (Table 1). In A, Cohen’s kappa was 0.946 between R1 and R2, with a 13% disagreement rate. In 56% of the disagreements, R3 agreed with R1. In B, Cohen’s kappa was 0.874 for R1 and d; R1 referred more than d. In A and B, DLAI referred more than R1, R2, and H/C. DLAI referred all eyes with > 1 VTD (1%) for further examination. Grading times for DLAI, R1, and R2 were 30, 129, and 68 seconds.
Conclusions :
DLAI performed best in DR and glaucoma detection; a potential solution for the high ungradable rate can be for DLAI to re-center uploaded images. DLAI can increase efficiency and accessibility of screenings for multiple VTDs, in both underserved populations and clinic. The ability to minimize direct contact confers an advantage during COVID-19. Further studies will investigate DLAI use in VTD progression.
This abstract was presented at the 2022 ARVO Annual Meeting, held in Denver, CO, May 1-4, 2022, and virtually.