Automatic Measurements of Smooth Pursuit Eye Movements by Video-Oculography and Deep Learning-Based Object Detection

Purpose The purpose of this study was to develop a technique that would combine video oculography (VOG) with single shot multibox detector (SSD) to accurately and quantitatively examine eye movements. Methods Eleven healthy volunteers (21.3 ± 0.9 years) participated in this study. Eye movements were recorded during the tracking of the target using a custom-made eye tracker based on EMR-9 (NAC Image Technology Inc.). The subjects were asked to fixate on the nose of the rabbit-like target (visual angle was 0.1°) that was manually moved to a distance of 1 meter by the examiner during the eye movement test. The test produced 500 images from the VOG external camera and these images were divided into 3 groups (300, 100, and 100) for training, verification, and testing. The performance of the SSD was evaluated with 75% average precision (AP75), and the relationship between the location of the fixation target (calculated by the SSD) and the positions of both eyes (recorded by the VOG) was analyzed. Results The AP75 of the SSD on one class of targets was 97.7%. The horizontal and vertical target locations significantly and positively correlated with the horizontal dominant (horizontal, adjusted R2 = 0.984, P < 0.001; vertical, adjusted R2 = 0.955, P < 0.001) and nondominant (horizontal, adjusted R2 = 0.983, P < 0.001; vertical, adjusted R2 = 0.964, P < 0.001) eye positions. Conclusions Our findings suggest that using VOG with SSD is suitable to evaluate eye version movements in the standard clinical assessment.