Abstract
Purpose :
Globally, Type 1 Diabetes Mellitus (T1DM) impacts 4.5% of the world population. Yearly eye exam to prevent complications in T1DM is recommended. Interpreting images with artificial intelligence (AI) can assist in accuracy and efficiency of image evaluation. The purpose was to assess AI performance for assisted tele-ocular screening (TOS) vs human reader during a large TOS.
Methods :
Use of AI (Visulytix Pegasus, UK) was investigated on 492 retinal images obtained from 246 participants ranging from 4 to 70 years old (Canon CR-2 non-mydriatic retinal camera (Tokyo, Japan)). Images were interpreted by a human reader (r1) to compare with AI, while a second reader (r2) viewed 109 images with discrepancies between AI and r1. Images were uploaded to a cloud server where AI analysis generated a grading report for each image. Reports included parameters on diabetic retinopathy (DR), micro-aneurysms (MA), Exudates (E), hemorrhages (H), and vertical cup to disc ratios. Data was compiled with analysis for accuracy of diagnosis and interpretation time. SigmaPlot (SYSTAT, CA, USA) was used to perform statistical analysis, student t test and ANOVA.
Results :
AI interpretation was possible on 488 images (4 images unreadable, 0.2% of total) while r1 interpreted 485 images (7 images unreadable, 1.4% of total). AI interpretation was achieved in 14 seconds per 2 images (Right and Left), while r1 needed an average of 82 seconds to analyze the same images (p<0.001). Detection of H, MA, and E in participants were compared between AI and r1 for accuracy. AI and r1 found E in 5.1% and 4.0% (p=.43), H in 4.3% and 9.6% (p=.0015), and MA in 30.1% and 7.4% (p<.0001) of images respectively. In images with disagreement between AI and r1, r2 disagreed with AI 86% of the time (of these images 65% were from subjects < age 30, and of those age <30: 89% had significant retinal reflections or shine).
Conclusions :
AI was successfully integrated into the TOS process. AI assisted TOS was faster in image interpretation than a human reader. AI detection of DR parameters in T1DM participants varied; AI performed well in recognition of E, however, there were discrepancies in H and MA detection, which may have been due to retinal shine found in under 30 year old cohort. Further development with AI will aim to resolve false positive readings and integrate it optimally with TOS.
This abstract was presented at the 2019 ARVO Annual Meeting, held in Vancouver, Canada, April 28 - May 2, 2019.