Deep-learning approach to detect childhood glaucoma based on periocular photograph

Abstract

Childhood glaucoma is one of the major causes of blindness in children, however, its diagnosis is of great challenge. The study aimed to demonstrate and evaluate the performance of a deep-learning (DL) model for detecting childhood glaucoma based on periocular photographs. Primary gaze photographs of children diagnosed with glaucoma with appearance features (corneal opacity, corneal enlargement, and/or globe enlargement) were retrospectively collected from the database of a single referral center. DL framework with the RepVGG architecture was used to automatically recognize childhood glaucoma from photographs. The average receiver operating characteristic curve (AUC) of fivefold cross-validation was 0.91. When the fivefold result was assembled, the DL model achieved an AUC of 0.95 with a sensitivity of 0.85 and specificity of 0.94. The DL model showed comparable accuracy to the pediatric ophthalmologists and glaucoma specialists in diagnosing childhood glaucoma (0.90 vs 0.81, p = 0.22, chi-square test), outperforming the average of human examiners in the detection rate of childhood glaucoma in cases without corneal opacity (72% vs. 34%, p = 0.038, chi-square test), with a bilateral corneal enlargement (100% vs. 67%, p = 0.03), and without skin lesions (87% vs. 64%, p = 0.02). Hence, this DL model is a promising tool for diagnosing missed childhood glaucoma cases.