Dynamic Accommodation and Vergence Measurement: A Purkinje Reflections and Machine Learning Approach

By Aryaman PattnayakReviewed by Susha Cheriyedath, M.Sc.Dec 12 2023

An article published in the journal Scientific Reports proposes a novel technique for real-time dynamic accommodation and vergence measurement using Purkinje reflections and machine learning.

Accurate measurement of eye accommodation and vergence is crucial for diagnosing vision abnormalities and assessing the performance of 3D displays. While eye movements can estimate vergence, accurately quantifying real-time accommodation changes remains challenging.

The ability to adjust the shape of the crystalline lens for focusing at varying distances, known as accommodation, is integral to normal vision. Simultaneously, the eyes converge through rotation to fixate on objects, a phenomenon termed vergence. Precise measurement of accommodation and vergence responses offers valuable insights into refractive errors, binocular vision defects, near-work effects, and treatment efficacy.

Although autorefractors are commonly used in clinical settings to assess accommodation by measuring retinal refraction, they fall short of directly quantifying lens shape and power changes during accommodation. Another complex scenario is the vergence-accommodation conflict (VAC), particularly in stereoscopic displays where the accommodation distance to the screen and the vergence distance to the perceived image may differ. This discrepancy underscores the need for accurate and reliable accommodation measurement methods.

The recent scientific endeavor, detailed in the paper, addresses these challenges by proposing a novel approach that combines the analysis of Purkinje images with advanced machine learning algorithms. This enhances our understanding of accommodation and vergence dynamics and provides a potential solution to the limitations posed by existing measurement techniques. The implications of this research extend beyond clinical applications, encompassing the realm of 3D displays where accurate accommodation assessment is crucial to mitigate issues associated with the vergence-accommodation conflict. As we delve deeper into these advancements, the potential for groundbreaking contributions to vision diagnostics and display technologies becomes increasingly evident.

The Study Details

The investigation centered on the dynamic monitoring of accommodation and vergence in the human eye, leveraging the unique characteristics of Purkinje images influenced by the shape of the accommodating lens and eye rotation. The authors designed an experimental setup comprising a controllable red-green-blue (RGB) light-emitting diode (LED) panel strategically positioned to induce controlled accommodation-vergence responses across various distances. Additionally, a Near-infrared (NIR) LED was employed to illuminate the participant's eye, while an NIR camera captured images of the pupil and Purkinje reflections P1, P3, and P4.

An adjustable eye model based on ZEMAX was incorporated to simulate expected Purkinje images for different accommodated states. The optimal camera-LED-eye configuration was identified through meticulous matching of simulations and experimental data, ensuring the capture of all reflections within the 4D to 1D accommodation range. The RGB LED panel was sequentially illuminated from far to near points, and continuous images were acquired using the NIR camera during changes in participant viewing distance and angle.

The captured images underwent a thorough analysis, focusing on extracting key parameters such as pupil center and radius along x and y, as well as the centroid coordinates of P1, P3, and P4 reflections. These features served as systematic inputs to a multi-layer perceptron (MLP) machine learning algorithm, which was trained and tested using two distinct methods: (1) utilizing subsets of individual subject data and (2) employing data from other subjects with a two-point calibration.

Findings

The study's results underscore the efficacy of Purkinje images and machine learning for user-specific and generalized accommodation convergence measurements. Specifically, vertical P3 to P4 coordinates exhibited a robust correlation with accommodation, while the P1 to P4 distance demonstrated a significant correlation with vergence. Analytical equations predicted accommodation with a Root Mean Square Error (RMSE) of 0.21D and vergence with an RMSE of 0.19°.

Implementing a subject-specific MLP model further improved accuracy, achieving an accommodation RMSE of 0.18D and a vergence RMSE of 0.16°. Even in Leave-One-Subject-Out (LOSO) cross-validation, the MLP model demonstrated notable precision, with errors in accommodation and vergence measuring at 0.32D and 0.32°, respectively, after two-point calibration. The demonstrated technique showcases reliable sub-0.5D accommodation accuracy, thus holding significant promise for ophthalmic diagnosis and stereo display characterization applications.

Future Outlook

The investigation outlined in this study introduces a straightforward yet highly accurate solution for real-time monitoring of dynamic accommodation and vergence responses, utilizing commonly available camera and illumination hardware. At the heart of this approach is the precise capture of specific Purkinje reflections, coupled with the application of machine learning algorithms to correlate image features with changes in lens shape and eye rotation.

With minimal customization required through individual calibration data, the study establishes that errors below 0.5D are deemed acceptable according to diagnostic standards. The potential for enhancing cross-subject model performance is identified, suggesting that including additional subjects and the imposition of controlled conditions could contribute to improved accuracy and reliability.

To broaden the scope and validation of the methodology, future directions may involve extending the application to binocular tracking. Additionally, exploring the technique's performance across different age groups, explicitly comparing pre-presbyopes with aged subjects, could offer valuable insights and further validation.

In summary, the demonstrated technique exhibits considerable promise as an assistive tool for essential eye examinations and advanced applications such as the development of customized virtual reality headsets equipped with auto-accommodative displays. The simplicity, accuracy, and potential for customization make this approach a compelling candidate for further exploration and implementation in diverse ophthalmic and technological settings.