Performance Comparison of Deep Learning Architectures for Dry Eye Detection

A method for detecting dry eye is proposed. In this approach, a sheet of paper printed with grid lines is shaped into a cylinder and attached to a ring light. The grid lines are reflected onto the pupil when the device is mounted on a camera lens and used to photograph the eye. In instances of dry eye, the tear film becomes unstable, resulting in distortions of the reflected grid lines. Such distortions can be identified using machine learning, thereby facilitating the diagnosis of dry eye. Throughout this study, high-resolution images were captured using a digital camera. Regions exhibiting visually apparent distortions were manually delineated and incorporated into a dataset. Subsequently, the efficacy of 13 distinct machine learning architectures, encompassing VGG16, VGG19, ResNet50, InceptionV3, MobileNet, and DenseNet121, was evaluated to ascertain their capacity for detecting dry eye.