A deep learning-based algorithm identifies glaucomatous discs using monoscopic fundus photographs

Purpose: To develop and test the performance of a deep learning-based algorithm for glaucomatous disc identification using monoscopic fundus photographs.

Design: Fundus photograph database study.

Participants: Four thousand three hundred ninety-four fundus photographs, including 3768 images from previous Sydney-based clinical studies and 626 images from publicly available online RIM-ONE and High-Resolution Fundus (HRF) databases with definitive diagnoses.

Methods: We merged all databases except the HRF database, and then partitioned the dataset into a training set (80% of all cases) and a testing set (20% of all cases). We used the HRF images as an additional testing set. We compared the performance of the artificial intelligence (AI) system against a panel of practicing ophthalmologists including glaucoma subspecialists from Australia, New Zealand, Canada, and the United Kingdom.

Main Outcome Measures: The sensitivity and specificity of the AI system in detecting glaucomatous optic discs.

Results: By using monoscopic fundus photographs, the AI system demonstrated a high accuracy rate in glaucomatous disc identification (92.7%; 95% confidence interval [CI], 91.2%–94.2%), achieving 89.3% sensitivity (95% CI, 86.8%–91.7%) and 97.1% specificity (95% CI, 96.1%–98.1%), with an area under the receiver operating characteristic curve of 0.97 (95% CI, 0.96–0.98). Using the independent online HRF database (30 images), the AI system again accomplished high accuracy, with 86.7% in both sensitivity and specificity (for ophthalmologists, 75.6% sensitivity and 77.8% specificity) and an area under the receiver operating characteristic curve of 0.89 (95% CI, 0.76–1.00).

Conclusions: This study demonstrated that a deep learning-based algorithm can identify glaucomatous discs at high accuracy level using monoscopic fundus images. Given that it is far easier to obtain monoscopic disc images than high-quality stereoscopic images, this study highlights the algorithm’s potential application in large population-based disease screening or telemedicine programs.