Longitudinal Screening for Diabetic Retinopathy in a Nationwide Screening Program: Comparing Deep Learning and Human Graders

Figure 1

Flow of patients from the first to the second screening. The number of patients in the cohorts of deep learning (DL) and trained human graders (HG) is compared at each point of the screening. The reference standard for these cases was based on an overread by retina specialists (Methods). Screen positive/negative indicates patients whom the DL or HG indicated as positive/negative. In this simulated setting, only patients who were confirmed by retina specialists to have STDR (i.e., true positives) were referred for treatment. The remaining patients were entered into the second screening. Dropout before the second screening included patients with missing data in either DL or HG or determined as ungradable by the reference standard during the second screening.