Abstract
Purpose :
evaluate the accuracy of an autonomous artificial intelligence (AI) system for holistic maculopathy screening during occupational checkups
Methods :
A retrospective study on an unclean and unprepared dataset of 5,918 images from a population of 2,839 people evaluated with non-mydriatic cameras during routine occupational health checkups. The images were obtained by a trained technician using handheld non-mydriatic cameras on the participant's office premises. The camera models employed were Optomed Aurora (field of view (FOV) 50 degrees, 88% of the dataset), Zeiss Visuscout 100 (FOV 40 degrees, 9 % of the dataset), and Optomed SmartScope M5 (FOV 40 degrees, 3% of the dataset). Image acquisition took around two minutes per patient. The ground truth of the dataset was evaluated per eye; 2 retina specialists graded each image; if the grading did not match, a 3rd one reviewed the image to break the tie. The specific pathologies considered for evaluation were diabetic retinopathy (DR) (more than mild DR. Age-related macular degeneration (AMD) (mild or worse, suspected glaucomatous optic neuropathy (GON), and Nevus. Images with possible signs of maculopathy that didn't match the described taxonomy were classified as Other.
Results :
the assembly of algorithms for evaluating abnormalities had an area under the curve (AUC) of 0.963 with a sensitivity of 0.929 and a specificity of 0.868. The algorithms individually obtained: AMD: AUC 0.98; Sensitivity 0.938; specificity 0.957. DR AUC 0.95; Sensitivity 0.811; specificity 0.948. GON AUC 0.8892; Sensitivity 0.536 specificity 0.957. Nevus AUC 0.931; Sensitivity 0.867; specificity 0.907.
Conclusions :
The holistic IA approach is comparable to human experts at simultaneous detection of DR, AMD, GON, and Nevus. The integration of pathology-specific algorithms allows for obtaining high sensitivities without sacrificing specificity. Deep learning may facilitate wider screenings of eye diseases and become quick and reliable support for ophthalmological experts.
This is a 2021 ARVO Annual Meeting abstract.