Valentina Sarao1,2, Daniele Veritti1, Paolo Lanzetta3,4. 1. Department of Medicine-Ophthalmology, University of Udine, Via Colugna 50, 33100, Udine, Italy. 2. Istituto Europeo di Microchirurgia Oculare-IEMO, Udine, Italy. 3. Department of Medicine-Ophthalmology, University of Udine, Via Colugna 50, 33100, Udine, Italy. paolo.lanzetta@uniud.it. 4. Istituto Europeo di Microchirurgia Oculare-IEMO, Udine, Italy. paolo.lanzetta@uniud.it.
Abstract
PURPOSE: In this study, we evaluated the diagnostic performance of an automated artificial intelligence-based diabetic retinopathy (DR) algorithm with two retinal imaging systems using two different technologies: a conventional flash fundus camera and a white LED confocal scanner. METHODS: On the same day, patients underwent dilated colour fundus photography using both a conventional flash fundus camera (TRC-NW8, Topcon Corporation, Tokyo, Japan) and a fully automated white LED confocal scanner (Eidon, Centervue, Padova, Italy). All images were analysed for DR severity both by retina specialists and the AI software EyeArt (Eyenuk Inc., Los Angeles, CA) and graded as referable DR (RDR) or not RDR. Sensitivity, specificity and the area under the curve (AUC) were computed. RESULTS: A series of 165 diabetic subjects (330 eyes) were enrolled. The automated algorithm achieved 90.8% sensitivity with 75.3% specificity on images acquired with the conventional fundus camera and 94.1% sensitivity with 86.8% specificity on images obtained from the white LED confocal scanner. The difference between AUC was 0.0737 (p = 0.0023). CONCLUSION: The automated image analysis software is well suited to work with different imaging technologies. It achieved a better diagnostic performance when the white LED confocal scanner is used. Further evaluation in the context of screening campaigns is needed.
PURPOSE: In this study, we evaluated the diagnostic performance of an automated artificial intelligence-based diabetic retinopathy (DR) algorithm with two retinal imaging systems using two different technologies: a conventional flash fundus camera and a white LED confocal scanner. METHODS: On the same day, patients underwent dilated colour fundus photography using both a conventional flash fundus camera (TRC-NW8, Topcon Corporation, Tokyo, Japan) and a fully automated white LED confocal scanner (Eidon, Centervue, Padova, Italy). All images were analysed for DR severity both by retina specialists and the AI software EyeArt (Eyenuk Inc., Los Angeles, CA) and graded as referable DR (RDR) or not RDR. Sensitivity, specificity and the area under the curve (AUC) were computed. RESULTS: A series of 165 diabetic subjects (330 eyes) were enrolled. The automated algorithm achieved 90.8% sensitivity with 75.3% specificity on images acquired with the conventional fundus camera and 94.1% sensitivity with 86.8% specificity on images obtained from the white LED confocal scanner. The difference between AUC was 0.0737 (p = 0.0023). CONCLUSION: The automated image analysis software is well suited to work with different imaging technologies. It achieved a better diagnostic performance when the white LED confocal scanner is used. Further evaluation in the context of screening campaigns is needed.
Entities:
Keywords:
Artificial intelligence; Conventional flash fundus camera; Deep learning; Diabetic retinopathy; Telemedicine; White LED confocal scanner