Judith Massmann¹, Alexander Lichtenstein², Francisco M. López¹, Bob Williams²; ¹Frankfurt Institute for Advanced Studies, Germany, ²Health Access LLC

Numerous visual impairments, including amblyopia, congenital glaucoma, and even retinoblastoma, can be detected in red-eye reflex images from young children. The Bruckner test is traditionally performed by ophthalmologists in clinical settings to analyze such visual disorders. Despite early intervention being crucial to mitigate the consequences of these impairments, parents are often unaware of the importance of this test. However, thanks to the recent technological advances in smartphones and artificial intelligence, it is now possible to recreate the Bruckner test without an ophthalmologist. We present KidsVisionCheck, an app that can perform vision screening from the comfort of a mobile device using red-eye reflex images taken at home. KidsVisionCheck relies on deep neural networks trained on children's data collected and labeled by ophthalmologists. With an accuracy of nearly 90% for detecting abnormalities, KidsVisionCheck provides a very good performance for initial screenings, which should subsequently be confirmed and turned into a full diagnosis by professionals. Additionally, we perform a comparative study of numerous model architectures, pre-processing techniques, and image augmentations. We find that combinations of multiple of these approaches result in more accurate and robust classifications. Furthermore, this analysis allows us to identify the optimal conditions for data collection, so that we can provide immediate feedback to the users in terms of lighting conditions, camera positioning, and more. In sum, we propose that KidsVisionCheck can provide first vision screenings at home and in schools for the early detection of visual impairments.