The Integration of Eye Tracking and Deep Learning for Objective Contrast Sensitivity Function Measurement on the Children with and without Vision Deficit

Abstract

Purpose : We have previously demonstrated the combination of the eye-tracking technique and the preferential looking for the testing of contrast sensitivity function and further validated its accuracy and stability on adults. Here we aim to establish an objective contrast sensitivity function (CSF) measurement based on an eye-tracking technique and deep learning paradigm, which can accurately assess CSF in the children participants depending on neither the cognitive ability nor the language ability.

Methods : We proposed a CSF deep learning model that learns the pre-defined CSF curves, the adaptive stimuli generation and the CSF curve updating paradigms from a large-scale CSF database (5444 cases) of normal subjects and patients with different ocular diseases. Two hundred twenty-seven children aged 2 to 6 years were recruited to complete monocular CSF measurements with uncorrected visions. Thirty-three participants were randomly selected to participate in a repeated test session to verify the repeatability of the measurement. The accuracy of the CSF measurement was verified by comparison of the correlation between cut-off spatial frequency and visual acuity. Test-retest repeatability of this method was also evaluated. The feasibility of using CSF for disease classification was assessed by training a support vector machine classifier based on the CSF characteristics of different subjects.

Results : Good test-retest variability was evident for the computerized CSF measurement (ICC_cut-off=0.907, ICC_aulscf=0.937). The cut-off spatial frequencies measured were strongly correlated with the visual acuities (Pearson’s r=0.932, P<0.001). The results of the classification prediction model also indicated that CSF data measured with this method had clinical significance for the classification and diagnosis of different visual diseases (AUC_normal=0.944, AUC_ametropia=0.886, AUC_{anisometropia} =0.724).

Conclusions : A computerized CSF measurement based on preferential looking combining eye-tracking technique with deep learning provided an objective, fast and accurate assessment for children without average cognitive or verbal ability in the future vision detection or screening scene.

This abstract was presented at the 2023 ARVO Annual Meeting, held in New Orleans, LA, April 23-27, 2023.