OPHTHALMOLOGY CONGRESS 2022

To overcome the artificial intelligence (AI) black box dilemma to diagnose glaucoma damage, we applied a new convolutional neural network (CNN) of TrueColor confocal fundus images. With human-in-the-loop (HITL) data annotations, this CNN architecture neural network is useful not only for diagnosing glaucoma but also for predicting and locating detailed signs of glaucoma fundus, including: B. Splinter hemorrhage, glaucomatous optic nerve atrophy, vertical glaucomatous coupling, peripapillary atrophy and retinal nerve fiber layer defect (RNFL). The training was performed on a carefully selected private dataset of 1,400 high-resolution confocal fundus images. Of these, 1,120 (80%) are dedicated to training and 280 (20%) are dedicated to testing. We used a specially trained object detection method based on You Only Look Once version 5 (YOLOv5) to pinpoint the underlying condition. The 26 predefined conditions were annotated by a team of humans (composed of two glaucoma specialists and two

To compare mydriatic and eye-steering ultrawide field imaging (UWFI) with standard non-mydriatic UWFI examination in detecting peripheral retinal lesions in a myopic patient's cross-sectional observational study. 220 eyes from 110 myopia patients with known peripheral retinal lesions in at least one eye were recruited for Goldman 3 mirror contact lens examination. Using Optomap UWFI (Daytona, Optos, UK), non-mydriatic standard and eye-steering UWFI images were taken in the center and eye-steered in the up, down, nose, and temporal gaze directions. Standard and eye steering mydriatic UWFIs were recorded with a central line of sight and four different peripheral lines of sight. We compared the sensitivities of detecting peripheral retinal lesions at various UWFI settings. 141 (64.09%) eyes had peripheral retinal lesions. Sensitivity to detect low to high peripheral lesions is 41.84% (95% CI 33.62% to 50.54%) with standard non-mydriatic UWFI settings and 52.48% (95% CI 44, 08% to 60.7