A Deep Learning Model for Glaucoma Detection Outperforms Conventional Metrics

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Glaucoma is a leading cause of irreversible blindness.
• Early detection of glaucomatous damage is crucial for preserving vision.
• Optical coherence tomography (OCT) provides detailed imaging of the optic nerve and retina.

Purpose of the Study:

• To evaluate a deep learning model (DLM) for detecting glaucomatous damage from OCT reports.
• To compare the DLM's performance against conventional diagnostic metrics.

Main Methods:

• A ResNet-50 model was fine-tuned to classify healthy versus glaucomatous OCT reports.
• The DLM was trained on 4932 healthy and 207 glaucomatous OCT reports.
• Performance was assessed on three independent datasets, including eyes with clear glaucoma and those with subtle arcuate defects.

Main Results:

• The DLM achieved 99.5% specificity on healthy eyes.
• Sensitivities were 100% for clearly glaucomatous eyes and 95.1% for eyes with subtle defects.
• The DLM demonstrated a superior area under the receiver operating characteristic curve (AUROC) of 0.999 compared to conventional metrics.

Conclusions:

• The DLM exhibits high specificity and sensitivity in detecting glaucomatous damage from OCT reports.
• The DLM significantly outperforms established conventional metrics in diagnostic accuracy.
• This DLM shows potential as a screening tool and diagnostic aid for clinicians.