Evaluation of a deep learning supported remote diagnosis model for identification of diabetic retinopathy using wide-field Optomap
Background: We test a deep learning (DL) supported remote diagnosis approach to detect diabetic retinopathy (DR) and other referable retinal pathologies using ultra-wide-field (UWF) Optomap. Methods: Prospective, non-randomized study involving diabetic patients seen at endocrinology clinics. Non-expert imagers were trained to obtain non-dilated images using UWF Primary. Images were graded by two retina specialists and classified as DR or incidental retinal findings. Cohen's kappa was used to test the agreement between the remote diagnosis and the gold standard exam. A novel DL model was trained to identify the presence or absence of referable pathology, and sensitivity, specificity and area under the receiver operator characteristics curve (AUROC) were used to assess its performance. Results: A total of 265 patients were enrolled, of which 241 patients were imaged (433 eyes). The mean age was 50±17 years, 45% of patients were female, 34% had a diagnosis of diabetes mellitus type 1, and 66% of type 2. The average Hemoglobin A1c was 8.8±2.3%, and 81% were on Insulin. Of the 433 images, 404 (93%) were gradable, 64 patients (27%) were referred to a retina specialist, and 46 (19%) were referred to comprehensive ophthalmologist for a referable retinal pathology on remote diagnosis. Cohen's kappa was 0.58, indicating moderate agreement. Our DL algorithm achieved an accuracy of 82.8% (95% CI: 80.3-85.2%), a sensitivity of 81.0% (95% CI: 78.5-83.6%), specificity of 73.5% (95% CI: 70.6-76.3%), and AUROC of 81.0% (95% CI: 78.5-83.6%). Conclusions: UWF Primary can be used in the non-ophthalmology setting to screen for referable retinal pathology and can be successfully supported by an automated algorithm for image classification.
