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

Area of Science:

• Ophthalmology
• Artificial Intelligence
• Medical Imaging

Background:

• Childhood glaucoma is a leading cause of pediatric blindness.
• Accurate and timely diagnosis of childhood glaucoma presents significant challenges.
• Early detection is crucial to prevent irreversible vision loss.

Purpose of the Study:

• To develop and assess a deep learning (DL) model for detecting childhood glaucoma using periocular photographs.
• To evaluate the performance of the DL model compared to human expert diagnoses.
• To identify specific cases where the DL model outperforms human examiners.

Main Methods:

• Retrospective collection of periocular photographs from children with diagnosed glaucoma.
• Utilized a deep learning framework with the RepVGG architecture for automated recognition.
• Employed fivefold cross-validation to evaluate model performance, measuring AUC, sensitivity, and specificity.

Main Results:

• The DL model achieved an overall AUC of 0.95.
• The model demonstrated high sensitivity (0.85) and specificity (0.94).
• The DL model outperformed human examiners in detecting glaucoma in challenging cases, such as those without corneal opacity or with bilateral corneal enlargement.

Conclusions:

• The developed DL model is a highly accurate tool for diagnosing childhood glaucoma.
• This AI-driven approach shows potential for identifying missed cases and improving diagnostic rates.
• The DL model offers a promising supplementary tool for pediatric ophthalmologists and glaucoma specialists.