Abstract
Purpose :
To assess the performance of DL employing 7F-CFP in automated identification of eyes with moderately severe and severe NPDR among patients with diabetes in a US primary care setting.
Methods :
Eyes of 37,358 patients with diabetes were analyzed using data, including images, collected between 1999 and 2016 (Source: Inoveon Corporation, Oklahoma City, OK). DR severity and the presence of clinically significant macular edema were assessed from 7F-CFP by professional graders at a centralized reading center, and graded using the Early Treatment Diabetic Retinopathy Study Diabetic Retinopathy Severity Scale (DRSS). Prevalence of moderately severe or severe NPDR (DRSS 47–53), considering the worst DRSS score at the patient level, was 2.2% in this cohort. This set was split into 80% for model training, 10% for tuning, and 10% for testing, for a total of 29,890, 3732, and 3736 patients with 1,430,046, 180,534, and 180,135 images, respectively. A DL Inception-v3 model with transfer learning was trained at the image level on all 7 fields of view (including stereoscopy) for being either DRSS 47–53 or not. Predictions were averaged over all fields of view to provide a prediction at the eye level. We report model performance metrics in terms of area under the receiver operating characteristic (AUROC) curve, specificity, sensitivity, and positive predictive value.
Results :
The best model was selected based on performance on the tuning set, as well as the optimal cutoff for specificity and sensitivity maximizing the Youden index. The model performed well on the testing set with an AUROC of 0.962 (95% CI, 0.956, 0.967), sensitivity of 0.942 (95% CI, 0.934, 0.950), specificity of 0.946 (95% CI, 0.945, 0.948), and positive predictive value of 0.281 (95% CI, 0.272, 0.290).
Conclusions :
DL can support automated identification of eyes with DRSS 47–53. The model presented here can optimize screening of patients at risk of disease progression for participation in preventive clinical trials as well in clinical practice. Future research will further refine this proof-of-concept algorithm, including validation on other independent diverse datasets and in a real-world setting.
This is a 2021 ARVO Annual Meeting abstract.