Eye tracker accuracy: quantitative evaluation of the invisible eye center location

Stephan Wyder, Philippe C. Cattin

Abstract

This paper presents a high-precision methodology for validating the 3D localization accuracy of eye trackers. By utilizing a physical eye phantom and micro-computed tomography ($μ$CT) imaging, the researchers established a sub-millimeter ground truth for the "invisible" center of the eye. The study demonstrates that specialized eye trackers can achieve a 3D accuracy of 0.68 mm, meeting the stringent requirements for non-invasive medical interventions such as proton radiotherapy.

Introduction

Most eye-tracking research focuses on the point of gaze—the location on a screen where a person is looking. While sufficient for consumer applications, medical interventions such as robotic surgery or proton radiotherapy for intraocular tumors require knowing the exact 3D location of the eye center (specifically the center of corneal curvature) within a room's coordinate system.

Schematic of the eye model showing the "invisible" center of corneal curvature and other geometric axes.

The primary challenge is that the eye center is an "intangible point" hidden behind the iris, making it impossible to place physical markers on human volunteers for verification. Furthermore, human subjects introduce noise through micro-movements and varying levels of cooperation, which prevents the isolation of hardware and algorithmic errors. This research addresses these issues by developing a high-precision physical phantom and a validation pipeline based on $μ$CT imaging.

Methodology

To isolate system errors, the researchers developed a "testing stage" approach consisting of four primary components:

  1. Hardware Setup: A 4-Degree of Freedom (4-DOF) stage was constructed using two linear stages, a goniometer, and a rotation stage. A handcrafted glass eye prosthesis was mounted on this assembly.
  2. Ground Truth Acquisition: To define the "invisible" center of the glass eye, 15 high-resolution $μ$CT scans were performed. This allowed for the visualization of internal phantom geometry with 25 $μ$m voxel precision.
  3. Kinematic Modeling: A mathematical kinematic chain was developed to describe the eye center's movement relative to stage adjustments. This model was calibrated using the $μ$CT data to calculate the "true" 3D position of the eye center for any mechanical setting.
  4. Validation Pipeline: A video-based stereo eye tracker estimated the eye center, and its output was compared against the $μ$CT-derived ground truth.
The physical testing stage hardware and the flowchart of the validation process.

Experimental Results

The evaluation utilized a bespoke dataset of 15 $μ$CT volumes and 20 distinct eye orientations. The system achieved a mean 3D accuracy of 0.68 mm (median 0.67 mm). Performance varied by axis, with the lowest error in the vertical ($y$) axis (0.09 mm) and the highest in the depth ($z$) axis (0.54 mm).

Angular precision was also high, with a mean relative angle error of 0.50°. The $μ$CT-calibrated kinematic model itself demonstrated an accuracy of 31 $μ$m, ensuring that the ground truth was an order of magnitude more precise than the eye-tracking system under test. These results provide the first rigorous proof that non-invasive video tracking can meet the sub-millimeter "gold standard" required for medical eye tumor treatments.

Discussion and Conclusion

The use of a physical phantom and $μ$CT imaging allowed the researchers to identify a systematic bias along the longitudinal axis of the tracker, suggesting that specific optical or algorithmic properties can be further calibrated.

However, limitations remain. Current $μ$CT segmentation is manual, and the use of a "best-fit sphere" for the cornea may introduce minor errors. Future work will explore more complex surface-fitting methods and the development of "bio-realistic" phantoms that account for biological complexities such as pupil size changes and slight eye deformation during movement. Despite these limitations, this study establishes a vital benchmark for the spatial accuracy of medical-grade eye tracking.

Summary updated

Abstract

We present a new method to evaluate the accuracy of an eye tracker-based eye localization system. Measuring the accuracy of an eye tracker’s primary intention, the estimated point of gaze, is usually done with volunteers and a set of fixation points used as ground truth. However, verifying the accuracy of the location estimate of a volunteer’s eye center in 3D space is not easily possible. This is because the eye center, the center of corneal curvature, is an intangible point.
BibTeX
@article{wyder2018eye,
  author = {Stephan Wyder and Philippe C. Cattin},
  title = {{Eye tracker accuracy: quantitative evaluation of the invisible eye center location}},
  journal = {Int. J. Comput. Assist. Radiol. Surg},
  year = {2018},
  doi = {10.1007/s11548-018-1808-5},
}
Figure