Development of a Novel Deep Learning-Based Gaze Estimation Method for Detecting Strabismus

TL;DR

A new deep learning method estimates eye alignment from video to detect strabismus, showing promising results in both clinical and control cases.

Contribution

A novel deep learning-based gaze estimation algorithm for quantifying strabismus angles using video input.

Findings

The algorithm estimated gaze deviations in case subjects that aligned with clinical diagnoses.

A strong correlation (Spearman's r=0.961-0.965) was found between left and right eye gaze angles in control subjects.

The method showed potential as a non-invasive and accessible tool for strabismus assessment.

Abstract

Background

This study reports the development and preliminary validation of a deep learning (DL)-based algorithm capable of quantitatively estimating ocular alignment, specifically the direction and angle of eye position, using a technique known as gaze estimation. The purpose is to evaluate this algorithm as a novel method for detecting and quantifying strabismus.

Methods

A gaze-estimation model based on DL was applied to video input of ocular positions. The model is trained on a set of computer-generated eye images synthesized using UnityEyes (Computer Vision Laboratory, ETH Zurich, Zurich, Switzerland). The algorithm outputs visualizations of the right and left eyes along with estimated gaze angles. Two cases were examined: one without prior ophthalmologic history and one with a known diagnosis of exotropia. Additionally, the algorithm was applied to 10 subjects without ophthalmologic abnormalities to assess the correlation between the gaze directions of the left and right eyes.

Results

A total of 12 subjects were included in the study: two case subjects and 10 control subjects without ophthalmologic abnormalities. In Case 1 (no ophthalmologic history), the estimated gaze deviation was 4.3 degrees in the right eye and -0.5 degrees in the left. In Case 2 (diagnosed exotropia), the estimated deviation was 0.7 degrees in the right and -10.1 degrees in the left, closely reflecting the clinical diagnosis. Among the 10 control subjects, a strong correlation was observed between the gaze angles of both eyes (Spearman's r=0.961-0.965).

Conclusion

The algorithm demonstrated potential for quantifying strabismus angles through video-based gaze estimation. This method may offer a practical, non-invasive, and accessible approach for strabismus assessment, pending further validation against established clinical standards. Future work will enhance accuracy by incorporating multi-device datasets.

Figures