P3.270: Deep Learning Model-Based Eye Movement Perimetry for Visual Field Assessment

Background: Conventional visual field testing consists of fixation of head position and gaze fixation as well as manual clicking to measure stimulus detection; these elements present limitations for pediatric patients and affects optimal evaluation. Eye movement perimetry (EMP) provides relief of head and gaze fixation and removes the need for manual clicking in visual field assessment. Previous explorations of EMP have relied on specialized eye tracking equipment as well as limited metrics and gross head or eye movements; this study proposes a novel application of deep learning that leverages gaze estimation model predictions to assess stimuli detection in the evaluation of visual field.

Methods: A mock visual field assessment consisting of a central fixation cross and various peripheral stimuli was generated and a sample of 25 healthy subjects were video recorded taking the visual field test via either webcam and smartphone camera. The L2CS-Net gaze estimation deep learning model from Roboflow Inference was utilized to assess test subject detection of stimuli in the recorded videos. The angular yaw and pitch estimations from the deep learning model were used to estimate gaze direction and approximate gaze positions of test subjects, and metrics such as deviation from central fixation upon stimulus presentation were calculated in addition to saccadic reaction time (SRT) through data smoothing and peak-finding algorithms.

Results: The average SRT for healthy test subjects was 0.3720 seconds and the average deviation from central fixation upon stimulus presentation was 0.3310 radians.

Conclusion: Deep learning-based EMP shows potential as a means for accessible, accurate, and objective visual field assessment, enabling test subjects to visually engage with the testing environment in a more natural manner and providing relief from and mitigating limitations of conventional standard automated perimetry (SAP). This has particular salience in pediatrics, where allowing pediatric patients to naturally engage with presented stimuli for visual field testing could enhance patient cooperation and improve quality of evaluation. Further work is warranted to explore the merit of deep learning-based EMP, particularly with respect to the evaluation of patients with visual field defects as well as comparison with SAP.