Effect of simulated cataract on the accuracy of artificial intelligence in detecting diabetic retinopathy in color fundus photos.

ABSTRACT

PURPOSE: Artificial intelligence (AI) is often trained on images without ocular co-morbidities, limiting its generalizability. This study aims to evaluate the accuracy of a convolutional neural network (CNN) applied to color fundus photos (CFPs) with simulated cataracts (SCs) in detecting diabetic retinopathy (DR). METHODS: A database of 3662 CFPs (from Asia Pacific Tele-Ophthalmology Society (APTOS) 2019) was used. Using transfer learning, a CNN was trained to classify the training images as either DR or non-DR. The CNN was then applied to classify the testing images after an SC was applied, using varying degrees of Gaussian blur. RESULTS: Accuracy without SC was 97.0%, sensitivity (Sn) 95.7%, specificity (Sp) 98.3%. For mild SC, accuracy was 93.1%, Sn 91.8%, Sp 94.3%. For moderate SC, accuracy was 62.8%, Sn 31.4%, Sp 95.2%. For severe SC, accuracy was 53.5%, Sn 11.8%, Sp 96.5%. CONCLUSION: SCs significantly impaired AI accuracy. To prepare AI for clinical use, cataracts and other real-world clinical challenges affecting image quality must be accounted for.