Prediction of progression of diabetic retinopathy using automated lesion detection of retinal hemorrhages

Abstract

Purpose : The extent of retinal hemorrhages, particular those located more peripherally, have been suggested to be an important predictor of progression to proliferative diabetic retinopathy (PDR). In this study, we evaluate whether automatically detected hemorrhages on ultra-wide field (UWF) pseudocolor images are predictive of DR progression over two years

Methods : Seventy-three patients (104 eyes) with various stages of DR were recruited from a tertiary care center in India, and UWF pseudocolor images were captured (Optos Daytona Plus). The “EyeReadUWF” algorithm (Eyenuk) was used to automatically detect retinal hemorrhages on stereographically projected images. The resultant annotation “masks”/layers showing the automatically detected hemorrhages and the location of the center of the optic nerve head (ONH; marked on ImageJ) were then analyzed using the Optos “lesion detection” tool to quantify the hemorrhage frequency, total and average hemorrhage area, and average lesion distance from the optic nerve head (ONH). Hemorrhage frequency, total surface area (mm²), and distance from the ONH (mm) were compared between eyes that progressed to PDR over the next 2 years and those that did not, in order to define which features were associated with risk. Odds ratios for each parameter were computed following regression analysis

Results : Of the 104 eyes, 13 were excluded (2 had no DR, 11 had PDR) at baseline. Of the remaining 91 eyes, the automated software could not detect the lesions in 5 eyes due to image quality. Among the final study cohort of 86 eyes, 30 progressed to PDR. The mean (range ± standard deviation, SD) of the study parameters are shown in Table 1. Univariate regression analysis revealed that a greater distance of retinal hemorrhages from the ONH was significantly associated with a higher risk of progression (odds ratio, OR = 1.37; 95% confidence intervals, CI = 1.00-1.88; p = 0.05, whereas the hemorrhage count, and total surface area of hemorrhages did not reach statistical significance (Table 2)

Conclusions : A greater distance from the ONH of automatically detected hemorrhages on UWF imaging appears to be predictive of progression to PDR, highlighting the importance of more peripheral retinopathy in the risk stratification of DR patients. Future studies including automatic detection of other DR features, may further enhance the predictive model

This abstract was presented at the 2024 ARVO Annual Meeting, held in Seattle, WA, May 5-9, 2024.

 

Descriptive analysis

 

Regression analysis