Dynamic accommodation measurement using purkinje reflections and machine learning

Dynamic and precise measurements of eye accommodation and vergence are important for vision research, near-eye displays (NEDs), and the diagnosis of certain visual disorders and neurological diseases. Existing biomedical devices have important limitations because they are bulky and cannot be used to accurately measure eye accommodation dynamically. From an engineering perspective, implementing accommodation and vergence measurements in NEDs, such as augmented reality (AR) or virtual reality (VR) glasses, helps to address issues like vergence-accommodation conflict (VAC). Previous work on NEDs using adaptive focus techniques or retinal projection has reported reduced symptoms of VAC, but validating the accuracy of these systems remains challenging with current methods. Consequently, finding a simple, portable, and head-mountable method for measuring accommodation and vergence is an active area of research. There are existing approaches in the literature for measuring eye accommodation and vergence using Purkinje reflections from various layers of the eye, as well as machine learning (ML) methods. However, these methods generally do not measure accommodation but rather measure gaze and do not use all Purkinje reflections, which limits their accuracy. In this thesis, we present a simple and efficient method for measuring accommodation and vergence using four Purkinje reflections and ML techniques, specifically multilayer perceptron (MLP) networks. We employed a ZEMAX-based eye model to simulate the positions of the Purkinje reflections at various focal distances. Despite the system's sensitivity to environmental factors, we successfully collected experimental data from nine subjects. We employed two analytical approaches to train our MLP models: subject-specific analysis using the data of individual subjects and leave-one-subject-out cross-validation (LOSO-CV) complemented by two-point calibration. To the best of our knowledge, this is the first successful implementation of the LOSO-CV method for measuring accommodation and vergence. Our system demonstrated high accuracy, predicting accommodation within 0.22 diopters (D) using subject-specific data and achieving 0.40 D accuracy with two-point calibration using data from other subjects across nine subjects.