Detection of visual impairments using CNN and red-eye reflex images

Amblyopia, also known as “lazy eye”, is the most common cause of visual impairment in children and affects 3-6% of the population. It is caused by incomplete development of vision due to factors such as refractive errors, strabismus and deprivation of visual signals. Treatment consists of correcting the underlying problem in the visual system, e.g. by correcting refractive errors, realigning the eyes or removing opacities coupled with “penalization therapy” where the better eye is blurred or occluded to force development of vision in the amblyopic eye. However, the effectiveness of treatment decreases with age and is poor after the age of six. Even with early intervention, complete restoration of normal vision is rare and some degree of stereopsis impairment usually persists. Photovision screening based on the Bruckner test, in which the red reflex at the back of the eye is examined, can indicate amblyogenic factors. Advances in smartphone technology with its high-resolution cameras and high computing power offer the possibility of earlier detection of amblyogenic factors by enabling early evaluation of children by parents or caregivers without the need to consult a professional. We present KidsVisionCheck, an app that is suitable for vision screening and enables parents to check their children on a regular basis for eye abnormalities. It is based on the model which uses a convolutional neural network (based on a ResNet model) to detect visual abnormalities from red-eye reflex images. We trained our model using data collected from children’s vision screenings and labeled by a trained ophthalmologist. As the result, we were able to achieve good performance in detecting abnormalities based on the red-eye reflex.