Abstract
BACKGROUND: Prominent epicanthus could not only diminish the eyes' aesthetics but may be deceptive for its typical appearance of pseudo-esotropia. This study aims to apply a deep learning model to characterize the periocular morphology for preliminary identification. METHODS: This prospective study consecutively included 300 subjects visiting the ophthalmology department in a tertiary referral hospital. Children aged 7-18 years with simple epicanthus or concomitant esotropia and healthy volunteers who were age- and gender-matched were eligible for inclusion. Multiple metrics were extracted automatically and manually from facial images to characterize the periocular morphology and binocular symmetry. The dice coefficient (Dice), intraclass correlation coefficient (ICC), and Bland-Altman biases were calculated to evaluate their consistency. The receiver operating characteristic (ROC) curve determined the cut-off values of symmetry indexes (SIs) for distinguishing concomitant esotropia subjects from epicanthus ones. RESULTS: The Dice for eyelid and cornea segmentation were 0.949 and 0.944, respectively. The ICCs of the two measurements ranged from 0.898 to 0.983. Biases ranged from 0.16 to 0.74 mm. The periocular morphology of epicanthus eyes was significantly different from the normal ones, including palpebral fissure width (21.41±1.53 vs. 24.45±1.82 mm; P<0.01), and palpebral fissure height (8.91±1.37 vs. 9.60±1.25 mm; P<0.01). The ROC analysis yielded an area under the curve of 0.971 [95% confidence interval (CI): 0.950-0.991] with SI for distinguishing esotropia subjects. Its optimal cut-off value was 1.296 with 0.920 sensitivity and 0.910 specificity. CONCLUSIONS: Our study established a standard deep learning system for characterizing the periocular morphology of epicanthus and esotropia eyes with great accuracy. This objective method could be generalized to other periocular morphological assessments for clinical care.